A market map, engineering boundaries and practical choices
Historical market snapshot: 25 September 2026. Product versions, availability, ownership, licensing and planned retirement dates below describe that snapshot, not current conditions. This edition separates two views: a navigable market landscape and the architecture used to assemble a real system. The 25 categories below are purchasing and engineering decisions, not 25 services every application must deploy.
The main map presents significant alternatives rather than reducing each category to a single preferred vendor. Each category explains its value, the reason for inclusion and the choice it supports. Product links point to supplier or project materials. The discovery index lists 362 entries, including alternatives, adjacent products and lifecycle exclusions; additional references appear in the main categories.
Coverage is uneven, and inclusion is not qualification. “Included” means relevant to the category. “Preferred starting point” is an editorial judgment for a workload, not independently measured superiority. Planned retirement, maintenance and preview remain explicit qualifications. This is a dated selection aid, not a continuously verified inventory or a certified list of safe products.
How to use this atlas in the library
People and AI Agents explains the organization of work; the Engineering Handbook sets production engineering principles. This atlas connects those principles to implementation choices. It is a reference for a specific workflow, not an approved stack or a replacement for either book.
Compare architectural alternatives before vendors: local isolated execution versus managed compute; a separate memory service versus a memory layer over a shared database; individual components versus a managed platform or a finished domain product. The choice is how much responsibility a team retains, not how many features a supplier lists. A stored conversation does not recover a business operation, and a sandbox does not authorize its external effects.
For any shortlisted tool, identify what it provides, what the team must still implement, and one failure it cannot prevent. Task meaning, outcome criteria and authority remain the team’s responsibility even when execution infrastructure is purchased. Apply the handbook’s controls in the actual deployment; do not assume that a feature name proves enforcement.
A team’s assessment status is separate from vendor availability. An interesting reference illustrates an approach; a candidate deserves a bounded workload trial; tested requires a dated version, configuration, scenario and result; in use requires an identified deployment and owner. Availability (waitlist, preview, generally available, maintenance or retirement) is another axis. No entry in this atlas alone establishes either testing or production adoption.
The category map
1. Foundation Models
Value and engineering role. The model supplies reasoning, language, coding and multimodal capability. A useful landscape must show both frontier APIs and credible enterprise/open-weight alternatives: they differ in deployment rights, latency, language coverage and tool behavior.
| Products / projects | Reason for inclusion and differences |
|---|---|
| OpenAI · Anthropic · Gemini | General reasoning, structured outputs and multimodal baselines; compare the specific model and API feature, not the company as a whole. |
| DeepSeek · xAI / Grok | Alternative reasoning/coding families. The linked supplier materials describe V4.1-Flash and Grok 4.7; their suitability depends on the workload and deployment constraints. |
| Mistral · Qwen · Alibaba Model Studio | Private/open-weight and managed deployment choices. Qwen is a model family; Model Studio is a platform for consuming models, not its synonym. |
| Cohere · AI21 | Enterprise-oriented model and knowledge workloads; preserve Command/Embed/Rerank and Jamba/Maestro as distinct offerings. |
Selection judgment. Use a frontier API as a baseline, then add the alternative that changes a real constraint: deployment, language, modality or cost. No single provider is a universal winner across these constraints.
2. Inference & Model Serving
Value and engineering role. This category turns model weights into an operated service: capacity, batching, scaling, deployment and recovery. Separate inference software from the provider taking responsibility for running it.
| Products / projects | Reason for inclusion and differences |
|---|---|
| Hugging Face Inference Endpoints | Managed deployment close to the model ecosystem, with multiple engines and custom containers. |
| Together AI · Fireworks · Baseten | Open-model infrastructure alternatives: managed inference and dedicated capacity; Baseten is particularly relevant for custom model code and customer-cloud topology. |
| Modal · Groq · Replicate | Different purchasing units: programmable GPU execution; a latency-oriented serving option; convenient model APIs and custom deployments. They are not interchangeable hosting SKUs. |
| vLLM · SGLang · llama.cpp · MLX LM | Engines for self-operated GPU or local inference. Control increases, but capacity, security and upgrades become the operator’s job. |
Selection judgment. Managed serving buys operating capacity; an open engine buys implementation control. Compare utilization and total operating cost, not GPU price against token price alone.
3. Model Gateways & Routing
Value and engineering role. A gateway centralizes model access, provider routing, fallback, budgets and telemetry. It becomes valuable when several applications would otherwise duplicate these controls.
| Products / projects | Reason for inclusion and differences |
|---|---|
| OpenRouter · LiteLLM | OpenRouter emphasizes managed access and provider routing; LiteLLM provides a proxy the team can operate and configure. |
| Cloudflare AI Gateway · Portkey | Gateway controls in an edge/cloud ecosystem or a dedicated AI control plane. Portkey’s acquisition by Palo Alto Networks closed on 29 May 2026. |
| Bifrost · TensorZero · Helicone | Bifrost and TensorZero preserve alternative gateway/optimization approaches. Helicone remains visible as a maintenance-mode product, not a default greenfield choice. |
Selection judgment. Choose by ownership, policy enforcement and failure behavior. A fallback must preserve data-handling restrictions; a common API shape does not guarantee identical model semantics.
4. Agent Frameworks & Harnesses
Value and engineering role. Frameworks provide programming primitives; harnesses package the strategy for planning, tools, context and delegation. The difference determines how much behavior the application owns.
| Products / projects | Reason for inclusion and differences |
|---|---|
| LangGraph · Pydantic AI · Mastra | Explicit state graphs, typed Python and TypeScript workflows respectively: choose the programming model that makes the application testable. |
| Vercel AI SDK · Google ADK · Microsoft Agent Framework | Important ecosystem choices: TypeScript model/tool integration, Google’s agent development kit, and .NET/Python agent/workflow development. |
| CrewAI · Deep Agents · Claude Agent SDK | CrewAI makes coordinated agents and flows explicit; Deep Agents and Claude Agent SDK provide more opinionated harness behavior. |
Selection judgment. Claude Agent SDK belongs here: embedding a harness in your process is not buying managed hosting. Multi-agent coordination is useful only when division of work improves the outcome.
5. Managed Agent Runtimes
Value and engineering role. Managed execution removes parts of deployment, session isolation, scaling and operations. It must be distinguished from a framework library and from a durable business workflow.
| Products / projects | Reason for inclusion and differences |
|---|---|
| OpenAI Agents API | Managed harness candidate for long-running tool/code tasks; the beta status recorded in this snapshot is a material qualification, not a blanket GA endorsement. |
| AgentCore · Microsoft Foundry · Gemini Enterprise Agent Platform | Cloud-native hosting and adjacent identity/context services. Existing IAM and operating skills can outweigh a marginal feature advantage elsewhere. |
| LangSmith Deployment | Code-first stateful deployment with cloud, hybrid and enterprise self-hosting paths. |
Selection judgment. Inspect who owns running state, what survives redeployment and what can be exported. Runtime success does not prove that a requested external action succeeded.
6. Low-code & Enterprise Agent Platforms
Value and engineering role. These platforms reduce assembly and give business or product teams a maintained authoring environment. Their value is the combined workflow, connector, deployment and governance experience.
| Products / projects | Reason for inclusion and differences |
|---|---|
| n8n · Dify · Langflow | Operational automation, integrated RAG/application building and visual component assembly are different centers of gravity; licensing and new feature status matter. |
| Kore.ai / Artemis · Stack AI · Mindset AI · Voiceflow | Enterprise agent platform; visual enterprise workflows; governed harness/control-plane positioning; conversational customer-experience design. These are alternatives to investigate, not a ranked shortlist. |
| Copilot Studio · Workato · Relevance AI | Existing workplace ecosystem, enterprise iPaaS skills, and business-built agent teams respectively. |
| Flowise | Retained for historical importance and existing installations; upstream archival/EOL excludes it from the maintained greenfield defaults. |
Selection judgment. Prefer a platform when authoring and integration speed dominate. Evaluate the exit path for knowledge, permission mappings and operational state, not merely the visual graph. Credentials need a secure migration or reauthorization procedure; portable configuration should reference secrets rather than contain them, and permissions must be revalidated in the destination.
7. Agent Tools, APIs & MCP Integration
Value and engineering role. An integration fabric absorbs OAuth, token refresh, provider quirks, actions, sync and triggers. MCP standardizes access but does not eliminate that work.
| Products / projects | Reason for inclusion and differences |
|---|---|
| Composio · Arcade · Nango | Broad action catalog; delegated user execution and governance; developer-owned integration functions and synchronization. |
| Paragon · Merge · n8n | Embedded integrations/action surfaces; normalized provider data and API access; reusable workflows exposed as tools. Preserve these distinctions from agent-native catalogs. |
| ToolHive · AgentCore Gateway · agentgateway · ContextForge | A separate control-plane choice: tool discovery, aggregation, traffic and policy across servers. It does not itself supply every SaaS integration. |
Selection judgment. Decide first who owns integration code and user credentials. A large catalog is valuable only if the required actions, tenant mapping and revocation behavior fit.
8. Agent Compute & Sandboxes
Value and engineering role. Generated code and long-lived workspaces need isolated compute, files, network policy and lifecycle control. This category is broader than browser automation.
| Products / projects | Reason for inclusion and differences |
|---|---|
| Docker Sandboxes | An additional reference for isolated coding-agent execution. Local sandboxes use a microVM boundary; workspace sharing and network policy determine exposed resources. Local and cloud environments have distinct controls and lifecycles, so assess the intended mode. |
| E2B · Daytona · Sprites | Code sandbox; managed composable environments; persistent workspace. Disk snapshots, memory pause and process continuation are different guarantees. |
| Blaxel · Baseten · Modal · AgentCore | Agent-oriented compute, custom deployment and cloud-integrated execution paths. Blaxel’s acquisition by Baseten changes ownership, not the need to inspect the product contract. |
Selection judgment. E2B is a useful code-sandbox baseline; persistent-workspace needs justify comparison with Daytona/Sprites. Daytona’s current core is private. None of these environments provides automatic rollback of an external API action.
Boundary to test. With Docker Sandboxes, inspect the security model: a directly mounted workspace is writable, and host-side MCP integrations are explicit trust connections. Isolation does not grant business permission, make broad credentials safe, or undo an already completed API action. Test allowed and denied access using the actual mounts, tools and network policy.
9. Browser & Computer Use
Value and engineering role. This category reaches software designed for people when APIs are missing. Separate browser infrastructure, controller SDKs and managed task execution: they remove different engineering work.
| Products / projects | Reason for inclusion and differences |
|---|---|
| Browserbase · Steel · Bright Data Agent Browser | Hosted browser sessions and execution infrastructure; Bright Data adds a web-access/unblocking specialization. |
| Stagehand · Browser Use · Skyvern | Developer browser SDK; agent-driven browser control; a more integrated portal-workflow approach. |
| Cua · Mobilerun | Desktop and mobile actuation extend the environment beyond web pages; specialized alternatives rather than substitutes for every browser stack. |
Selection judgment. Buy sessions when you own the controller; buy a managed workflow when portal completion is the job. In either case, verify the resulting application state after timeouts or clicks.
10. Search, Web Retrieval & Scraping
Value and engineering role. External context has several jobs: discover a source, retrieve its content, maintain a corpus or produce a researched answer. Combining them under “scraping” hides meaningful product differences.
| Products / projects | Reason for inclusion and differences |
|---|---|
| Exa · Tavily · Perplexity API | Search/retrieval primitives and research-oriented APIs. Perplexity exposes search and agent surfaces; it is not merely a page scraper. |
| Firecrawl · Bright Data · Browse AI | Developer-oriented crawl/extraction; web access and collection infrastructure; no-code extraction and change monitoring. |
Selection judgment. Exa/Tavily fit source discovery; Firecrawl fits building a corpus; Browse AI fits operational monitoring without a custom crawler. Citation quality, freshness and reproducibility matter more than a single search demo.
11. Document Intelligence & RAG Ingestion
Value and engineering role. This layer converts messy files into structured evidence: layout, tables, fields, chunks and provenance. Retrieval cannot reliably recover information that parsing has already damaged.
| Products / projects | Reason for inclusion and differences |
|---|---|
| Docling · Reducto | Local structured parsing baseline versus managed extraction focused on fields and tables. |
| Unstructured · LlamaIndex / LlamaCloud | Broad ingestion/partitioning and managed parsing/indexing/retrieval workflows. LlamaParse is a product within the latter ecosystem, not the whole framework. |
| Firecrawl · Marker · MinerU | Web-origin ingestion and additional document-parser alternatives with distinct deployment and model-license boundaries. |
Selection judgment. Choose by document structure and review burden: exact fields, tables, reading order and evidence coordinates. A large supported-format list does not establish accuracy on the actual documents.
12. Vector & Knowledge Infrastructure
Value and engineering role. Knowledge infrastructure owns indexes, filtering, retrieval and operational scale. The choice is partly a data-architecture decision, not a contest between interchangeable vector databases.
| Products / projects | Reason for inclusion and differences |
|---|---|
| SurrealDB | A multi-model database alternative to a separate vector-only store. Assess shared graph, vector and structured records together with transactional and access-control requirements. The database is distinct from the Agent Memory application layer described below. |
| Pinecone · Weaviate · Qdrant · Milvus / Zilliz | Managed knowledge/search service, integrated retrieval platform, dedicated vector/hybrid service, and an open distributed engine with a managed ecosystem. |
| pgvector · Supabase · Elasticsearch | Relational proximity; a PostgreSQL/application platform with vector support; lexical/hybrid search. Supabase is not a pure vector vendor. |
| Voyage AI · Cohere · Jina AI | Embedding and reranking models influence retrieval quality independently of the chosen store; changing embeddings may require reindexing. |
Selection judgment. Start near existing data where that meets filtering and latency requirements. Dedicated retrieval is justified by relevance, scale or operational separation—not a universal document-count threshold.
13. Agent Memory & Context
Value and engineering role. Memory supports continuity across tasks and sessions: preferences, prior work and changing relationships. Its value depends on correction, expiry and provenance, not simply retaining more conversation.
| Products / projects | Reason for inclusion and differences |
|---|---|
| SurrealDB Agent Memory | A memory and knowledge application layer in front of SurrealDB, with provenance and temporal history on durable records. Useful for comparing a shared database-backed approach with separate memory services. The public access page offers a waitlist/design-partner application; confirm access and deployment terms rather than infer general availability from documentation. |
| Letta · Mem0 | Stateful agent design and memory management versus reusable user/session context. Letta is broader than a memory-only plugin. |
| Zep · Graphiti · Cognee | Managed temporal context, its distinct graph implementation, and graph-oriented knowledge/memory pipelines. Shared concepts do not imply identical deployment rights. |
| Hindsight · Hyperspell | Specialist alternatives for testing different memory and context-management approaches. |
Selection judgment. Add memory when repeated tasks improve measurably. Keep mutable authoritative records outside inferred memory; deleting source data must also address derived facts. Retired Letta V1 or Zep CE does not mean the whole current product is retired.
Boundary to test. A common database does not make extracted beliefs authoritative or grant every agent access. Verify source attribution, correction, retention/deletion of derived records and authorization at retrieval time. History preservation and a required deletion policy must be reconciled explicitly. The database’s transactions do not make actions across external tools atomic.
14. Voice & Realtime Agents
Value and engineering role. Realtime agents combine transport, turn-taking, speech, tools and handoff. Buying speech synthesis alone does not remove the work of operating a conversation.
| Products / projects | Reason for inclusion and differences |
|---|---|
| ElevenLabs / ElevenAgents · Retell · Vapi | Integrated conversational agent tooling and managed call operations; compare supported channels, telephony and transfer behavior per product. |
| LiveKit Agents · Pipecat | Programmable realtime sessions/media and composable multimodal pipelines when engineering control is important. |
| Deepgram · Cartesia | Speech and realtime building blocks; Deepgram also exposes a Voice Agent API. These are not all the same kind of full contact-center platform. |
Selection judgment. Use managed services when telephone operations dominate; use a programmable media stack for custom interaction. Measure interruptions, tool latency and human transfer alongside voice quality.
15. Evals, Tracing & Observability
Value and engineering role. This layer links a production failure to a dataset case, an experiment and a release decision. It is the feedback loop for agent engineering, not just logging.
| Products / projects | Reason for inclusion and differences |
|---|---|
| Langfuse · Phoenix · Braintrust · LangSmith | Framework-neutral/private tracing; OTel/OpenInference debugging; evaluation-centered collaboration; an integrated agent-engineering ecosystem. |
| W&B Weave · Galileo · Opik · MLflow | Experiment/evaluation workflows in the W&B ecosystem; quality monitoring; optimization with tracing; continuity with existing ML operations. |
| Promptfoo · LangWatch · Scenario | Regression/adversarial checks and multi-turn simulation complement the trace store. |
Selection judgment. Choose one primary quality loop and add specialist testing where needed. Export traces, datasets and evaluator lineage together, subject to access and retention controls. Phoenix identifies ELv2 as its repository license. Check that license for the intended use and separately check the terms of any commercial offering; self-hosting is not by itself evidence that a paid license is required.
16. Runtime AI Security
Value and engineering role. Runtime security detects and constrains prompt injection, tool abuse, leakage and other AI-specific attacks. Supply-chain scanning and attack simulation extend protection before deployment.
| Products / projects | Reason for inclusion and differences |
|---|---|
| Check Point / Lakera · Prisma AIRS · HiddenLayer | Inline inspection and broader AI security suites; HiddenLayer explicitly covers discovery, supply chain, simulation and runtime security. |
| Protecto · NeMo Guardrails | Sensitive-data detection/masking and programmable input/retrieval/tool/output controls. Protecto is distinct from Protect AI. |
Selection judgment. Buy according to the failure mode and enforcement point. A detector is not the business authorization system; assess false positives, bypass paths and behavior if the security service is unavailable.
17. Governance & AI GRC
Value and engineering role. Governance establishes what AI systems exist, who owns them, which risks are accepted and what evidence supports oversight. It is a different buying decision from inline attack detection.
| Products / projects | Reason for inclusion and differences |
|---|---|
| Credo AI | A dedicated governance platform covering inventory, risk workflows and evidence assembly. |
| Fiddler AI · Arize | Observability, evaluation and control-plane evidence relevant to oversight. They overlap governance without becoming identical to a dedicated GRC system. |
Selection judgment. Select governance when portfolio ownership and evidence are the bottleneck. A dashboard or policy inventory does not itself block an unauthorized tool call.
18. Identity & Agent Authorization
Value and engineering role. Agents need attributable identities, bounded delegation, consent and revocation. Authentication, credential custody and the decision to allow a particular action remain separate responsibilities.
| Products / projects | Reason for inclusion and differences |
|---|---|
| Stytch · Descope · WorkOS AuthKit · Auth0 | Connected-app/MCP OAuth, agent identity integration, an OAuth server for MCP and delegated credential access in an established identity stack. |
| Arcade · Composio | Tool-level delegated authorization and credential execution paths complement the application’s identity provider. |
| SPIRE · OPA · Cerbos · OpenBao | Workload identity, policy decisions, resource/action authorization and secrets lifecycle respectively. |
Selection judgment. Preserve the existing identity system where possible. Bind approval to the exact action and enforce it at execution time; an OAuth scope is not a complete domain permission model.
Database enforcement example. SurrealDB permissions distinguish record users from system users. Table and field PERMISSIONS apply to record users (and enabled guests); system users use roles instead and are not restricted by those clauses. An agent-specific system login is therefore not proof of row isolation. Test the effective identity on every access path, including retrieval. Row restrictions and field visibility are separate controls; a metadata filter or broad service credential is not a substitute.
19. Agentic Payments & Commerce
Value and engineering role. This optional category enables delegated purchasing, paid tools and agent-accessible commerce. Separate order state, authority to spend, payment, delivery and reconciliation.
| Products / projects | Reason for inclusion and differences |
|---|---|
| Stripe · Shopify · AgentCore payments | Merchant acceptance, catalog/checkout/order ownership and managed paid-resource purchasing within AWS. |
| Crossmint · Coinbase Agentic Wallet · Privy · Nevermined · Skyfire | Scoped credentials and checkout; developer wallets; programmable signing; metering/plans; identity-linked charges. Different components of a transaction. |
| x402 · MPP · AP2 · ACP · UCP | Protocols for paid resources, mandates and commerce; protocols are not wallet providers or payment processors. |
Selection judgment. Add only the missing role. MPP was co-authored by Stripe and Tempo; its Payment HTTP authentication scheme is an individual Internet-Draft, not an RFC; agent features may be preview even within a mature payment provider. A payment receipt does not prove delivery.
November 2026 reading companion. For a transaction-level account of authorization, payment, delivery and recovery, see When an AI Agent Can Pay, Who Owns the Purchase?. This contextual essay does not update the dated landscape snapshot or certify any payment provider.
20. Agent Protocols & Interoperability
Value and engineering role. Protocols reduce repeated integration work across ownership boundaries. The contract includes revision, transport, authentication, extensions, cancellation and error behavior—not just a protocol logo.
| Products / projects | Reason for inclusion and differences |
|---|---|
| MCP · A2A · AG-UI | Tools/context access; remote-agent task lifecycle; agent–application interaction events. These are complementary boundaries. |
| A2UI · MCP Apps · OpenTelemetry | Declarative UI, embedded tool interfaces and telemetry conventions; each requires matching host/renderer/version support. |
| Agent Skills · AGENTS.md · FastMCP | Instruction/resource packaging, repository guidance and a Python MCP implementation. Specifications and implementations must remain distinct. |
Selection judgment. Adopt a protocol at a real integration boundary. MCP spec defaults and SDK behavior can differ; A2A task completion does not guarantee exactly-once external effects.
21. Durable Execution & Business Workflows
Value and engineering role. Durability preserves task identity and progress through crashes, human waits, retries and deployments. It deserves a visible category because a conversation checkpoint is not a recovered business operation.
| Products / projects | Reason for inclusion and differences |
|---|---|
| Temporal · Restate · DBOS | Cross-service workflow history and recovery; durable service handlers/entity serialization; PostgreSQL-oriented durable application execution. |
| LangGraph | Graph checkpoints and resumable reasoning are adjacent capabilities; inspect the outer operation and external side effects separately. |
Selection judgment. Temporal is one starting point for evaluating long-running cross-service work; Restate/DBOS offer different operating models. Test a successful external commit followed by a lost acknowledgment before accepting retry behavior.
22. Human Interfaces, Generative UI & Approvals
Value and engineering role. The interface makes work inspectable: progress, evidence, artifacts, edits, consent and cancellation. This is broader than a chat box and separate from voice infrastructure.
| Products / projects | Reason for inclusion and differences |
|---|---|
| AI SDK UI · assistant-ui · CopilotKit · ChatKit | Custom streaming primitives; reusable chat components; shared application state; embedded chat. ChatKit remains independent of Agent Builder retirement. |
| Tambo · OpenUI · json-render | Generated interfaces from controlled components/schemas. They reduce UI assembly but do not grant tool permissions. |
Selection judgment. Use stable review surfaces for consequential actions and generated UI for exploration. Bind approval to immutable intent; closing the window or losing the stream may not stop server work.
23. Model Adaptation & Data Preparation
Value and engineering role. Adaptation changes model behavior when prompting and retrieval cannot close a measured gap. Data preparation, labels and reward validity often dominate the quality of the result.
| Products / projects | Reason for inclusion and differences |
|---|---|
| Tinker · PEFT · TRL · verl | Managed compute behind a custom training loop; parameter-efficient adaptation; post-training/RL tooling. These are different levels of training control. |
| Argilla · Label Studio · Distilabel · NeMo Curator | Human feedback/annotation and synthetic-data/curation pipelines address the input side of adaptation. |
Selection judgment. Do not train merely because the option exists. Require held-out improvement and deployment rights for the resulting model. Human-data workforce economics and reward design need a separate assessment; this map does not cover them in depth.
24. Enterprise Context Sync & Permissions
Value and engineering role. An enterprise assistant needs source updates and permissions to travel with data. This category sits between connectors and retrieval: an excellent index still leaks or misleads if synchronization is wrong.
| Products / projects | Reason for inclusion and differences |
|---|---|
| Airweave · Airbyte · Debezium | Context delivery through sync/federation, general data movement/CDC and database change streams. Their existence does not prove end-to-end ACL propagation. |
| Elasticsearch · LlamaIndex / LlamaCloud | Search/ingestion connector paths belong here as secondary roles; maturity and deletion guarantees must be assessed per connector. |
Selection judgment. Design revocation and deletion before adding more sources. Authoritative identity must drive filtering; tenant or user names inside a prompt are not access controls.
25. Applied Agent Systems
Value and engineering role. Applied systems sell completed domain work, not reusable infrastructure. Keep them visible but separate: the build-versus-buy decision is about accepted outcomes, exceptions and human rework.
| Products / projects | Reason for inclusion and differences |
|---|---|
| Devin · Fin · Sierra · Decagon | Delegated engineering toward reviewable changes; customer-service systems with different procedures and channels. |
| Glean Deep Research · Claygent · incident.io Investigations · IncidentFox | Enterprise research, GTM research, incident investigation and an open investigation alternative. Investigation is not unrestricted remediation. |
| Vic.ai · Numeric · HighRadius · Modern Treasury AI | Accounts payable, close/accounting, cash application/collections and payment operations. New AI modules retain their own maturity status; Modern Treasury AI remains a qualified watchlist case. |
Selection judgment. A useful applied product can replace more engineering work than a platform. Evaluate its exact job boundary and definition of a billable outcome; a mature suite does not prove every new agent feature mature.
Boundaries that matter
- Framework, harness, runtime and durability: libraries provide primitives; harnesses organize planning and tools; runtimes execute processes; durable engines preserve progress across failures and waits. Session storage alone does not establish durable business execution. LangGraph persistence, Temporal.
- Integration and MCP: a standard tool interface does not replace provider-specific authentication, token refresh, rate limits, synchronization or tenant isolation. Nango, MCP specification.
- Retrieval, memory and authoritative records: indexes and memory are derived representations. Permissions, freshness, correction and deletion must follow the source lifecycle. Current operational facts belong in authoritative systems. pgvector, Elastic connector permissions.
- Guardrails and authorization: detecting suspicious content does not grant authority or enforce a business policy. Check permissions at the point of action, against the actual payload and current state. OPA, SPIRE.
- Telemetry and outcome evidence: traces explain system behavior; evaluation measures quality; an authoritative result confirms what actually happened. Critical invariants need deterministic checks. OpenTelemetry GenAI.
RAG and multi-agent coordination are architectural patterns. Low-code is a development approach. Open source is a property of a particular component and license. None replaces these responsibility boundaries.
Choosing a practical starting point
| Need | Starting point | Decisive test |
|---|---|---|
| A first agent-enabled product | One model adapter, explicit workflow, a small typed tool set and a quality loop | End-to-end task completion on representative cases |
| A long-running operation | Compare Temporal with Restate/DBOS | Recovery after an external action succeeds but its acknowledgment is lost |
| Private inference | Compare vLLM with SGLang | Quality, latency, recovery and full operating cost on the target model |
| Document-heavy work | Docling or Reducto; existing PostgreSQL/pgvector before adding dedicated search | Field accuracy, retrieval quality, permissions and deletion |
| A custom application interface | AI SDK UI or assistant-ui; CopilotKit for shared state | Approval bound to the exact action, cancellation and reconnect |
| Voice workflows | LiveKit Agents/Pipecat or managed Retell/Vapi | Interruptions, tool latency, transfer and recovery |
| An existing cloud estate | Evaluate its integrated agent platform first | Identity fit, operational ownership and export of one real workflow |
Use the same ten-part scorecard for every candidate: architecture; developer experience; production readiness; open source/self-hosting; interoperability; security; observability; adoption/momentum; pricing/lock-in; category durability. Record evidence and unknowns separately. Do not average unrelated workloads into a universal vendor score.
Measure cost per correct completed outcome, including retries, tools, compute, human review, rework and operations. Token price alone is insufficient.
Three useful platform shapes
Knowledge product: begin with parsing only where needed, permission-aware retrieval, citations, a user interface and a trace/evaluation loop. This architecture is sufficient for many research and support applications. Add memory for useful continuity; add autonomous action only when the product actually needs to change another system.
Operational agent: put a bounded harness or model step inside an explicit workflow, connect typed tools through controlled credentials, and verify the resulting state. Add durable execution for waits and recovery. This pattern fits support escalation, provisioning, document processing and other processes whose allowed stages are known.
Shared engineering platform: consolidate model access, tool discovery, identity/policy, deployment and quality services only after several applications repeat these needs. Keep domain workflows with their owning teams. A universal agent builder is not a prerequisite for a useful first application.
Applying the map
The clearest default is a small composable application or an existing-cloud platform, selected by ownership and operating constraints. Most teams do not need every layer as a separate purchase. A framework does not automatically require a separate harness, memory product, vector database and multi-agent coordinator.
Use two or three contrasting candidates for the unresolved decision: typed code versus explicit graphs; existing database versus dedicated search; operated framework versus managed service. The selection logic above already narrows the field. The remaining empirical check should establish task quality, failure recovery and cost under the organization’s actual conditions, rather than restart market discovery.
Use the ten assessment dimensions above as different questions. Architecture and interoperability inform integration fit; production readiness and security inform operating constraints; DX and observability affect engineering effort; self-hosting, pricing and lock-in shape control and exit cost; adoption and category durability inform continuity risk. A documented capability can be relevant while its performance or commercial terms remain unknown.
Lifecycle and convergence
Lifecycle can change the shortlist. At the snapshot date, Flowise upstream is archived, Daytona’s current core is private, and GraphRAG and TGI are in maintenance-oriented modes. These statements apply to the specified components, not every service from their owners. Flowise notice, Daytona, GraphRAG, TGI.
Agent Builder is deprecated with shutdown planned for 30 November 2026; this does not extend to ChatKit. Tavily’s acquisition by Nebius closed on 19 February 2026. Announcements, closing dates, renames and future retirements must be recorded separately. OpenAI deprecations, ChatKit, Nebius disclosure.
Execution and inference, quality and observability, and security and governance are converging into broader suites. Procurement may consolidate; responsibility boundaries remain. The durable platform capabilities are clear domain contracts, reliable execution, enforceable authority, integration lifecycle and verifiable outcomes.
Material lifecycle and classification updates
Portkey’s acquisition by Palo Alto Networks closed on 29 May 2026; gateway and security responsibilities still need separate evaluation. Closing announcement. Helicone remains in maintenance following its move to Mintlify. Maintainer notice.
Stytch joined Twilio through a completed acquisition on 14 November 2025. Acquirer confirmation.
The model references include DeepSeek V4.1-Flash and Grok 4.7. The latter documentation carries SpaceXAI branding; “xAI / Grok” is retained here for recognition. These links identify supplier descriptions, not independent performance comparisons. DeepSeek release, Grok documentation.
Category placement is deliberate: Claude Agent SDK is an embedded harness; Supabase is an application/data platform with vector support; Protecto is distinct from Protect AI; Fiddler and Arize overlap governance through observability/control evidence. Shared ownership or a broad marketing label does not merge these contracts.
Evidence and scope
The category descriptions synthesize public supplier documentation, project repositories, protocol specifications and lifecycle notices. Links beside product-specific statements identify those sources; the category boundaries and selection judgments are editorial synthesis. Supplier claims are not independent verification of performance, security or business outcomes.
Docker Sandboxes, SurrealDB and SurrealDB Agent Memory illustrate architectural alternatives. Throughout the map, a documented capability must be distinguished from demonstrated suitability: inclusion does not establish benchmark performance, deployment qualification, security assurance or negotiated pricing.
Coverage varies by entry and is not a market census. This edition does not provide independently reproducible vendor scores. Hardware markets, robotics and training-data workforce operations are outside its detailed scope. Before adoption, recheck the exact product version, availability, license and operating terms, then test the relevant failure modes in the intended deployment.
Full discovery index
This index lists 362 entries, not 362 qualified vendors. Names may refer to a model version, library, product family, protocol or platform. It is a historical discovery record rather than a current active-vendor shortlist. Lifecycle qualifications in the main text take precedence over an old label. Docker Sandboxes and the SurrealDB references appear in the main map rather than this index.
Use the main categories for selection logic and the index to locate alternatives. An entry identifies a source and an architectural role; it is not a rating, proof of suitability, benchmark result or deployment qualification.
| Entity / supplier or project source | Role |
|---|---|
| 11x — Digital workers | sales revenue |
| A2A | interface contract |
| A2A Python SDK | remote agent SDK |
| A2A TCK | qualification tooling |
| A2UI | interface contract |
| A2UI renderers | native UI renderer |
| Abridge — Clinical documentation | healthcare |
| ACP — Agent Client Protocol | interface contract |
| ACP — Agent Communication Protocol | interface contract |
| ACP — Agentic Commerce Protocol | interface contract |
| Activepieces | low code automation |
| Aembit — Agentic/workload IAM | workload identity |
| AG-UI | interface contract |
| AG-UI SDK | frontend transport SDK |
| Agent Skills | interface contract |
| Agenta | prompt/eval collaboration |
| agentgateway | tool control plane |
| AgentLedger maintainers — AgentLedger | execution evidence runtime |
| AgentOps | agent observability |
| AgentPay (agentpayement.com) — Public Alpha Sandbox | agent wallet orchestration |
| AGENTS.md | interface contract |
| AgentWallet (agentwallet.ai) — Wallet/card platform | agent wallet card |
| Agno SDK / AgentOS | integrated agent platform |
| Airbyte — Database CDC | enterprise data cdc |
| Airweave | enterprise context connectors |
| Amazon Bedrock AgentCore | hyperscaler |
| Amazon Bedrock AgentCore Gateway | tool control plane |
| Amazon Bedrock AgentCore payments | payment orchestration |
| Amazon Bedrock AgentCore Runtime | managed compute runtime |
| ANP | interface contract |
| Anthropic — Claude Agent SDK | agent harness |
| Anthropic — Claude API / Opus 5.5 | frontier model api |
| Anthropic — Claude Code | software engineering |
| AP2 | interface contract |
| Arcade | delegated tool execution |
| Argilla | training eval data curation |
| Argilla — Distilabel | synthetic training eval data |
| Arize Phoenix | trace/eval platform |
| Artisan — Ava | sales revenue |
| assistant-ui | chat presentation |
| Astrix — AI Agent Discovery | nhi inventory |
| Atlar — Treasury agents | finance treasury |
| Auth0 — AI Identity / Token Vault | delegated authorization |
| AuthZed — Authorization platform | relationship authorization |
| Axolotl | adaptation tooling |
| Baseten — Inference Platform | managed inference |
| BerriAI — LiteLLM Proxy | model gateway |
| BlackLine — Financial operations platform | finance close |
| Bland | managed voice platform |
| Blaxel | executable environment |
| Bolna | specialist voice orchestration |
| Braintrust | evaluation/data platform |
| Browser Use — browser-use | browser agent sdk |
| Browserbase Browser API | cloud browser |
| C1 | generative UI predecessor |
| Cartesia speech APIs | speech model infrastructure |
| Catena Labs / ACK — Agent Commerce Kit | protocol reference components |
| Cedar policy engine | policy decision |
| Cekura | voice/conversation testing |
| Cerbos — PDP + optional Hub | policy decision |
| Cerebras — Inference API | specialized inference |
| Chainlit maintainers — Chainlit | Python conversational app framework |
| Check Point AI Guardrails | runtime ai security |
| Chutes | managed inference |
| Circle — Agent Stack | wallet and marketplace |
| Claude Managed Agents | managed harness runtime |
| Claygent | sales revenue |
| Cloudflare — Agents SDK | stateful cloud runtime |
| Cloudflare — AI Gateway | model gateway |
| Cloudflare — MCP Server Portals | tool control plane |
| Cloudflare — Monetization Gateway | edge monetization |
| Cognee | agent memory |
| Cohere — Command / Embed / Rerank | specialist models |
| Cohere — Rerank | reranking |
| Coinbase Agentic Wallet | agent wallet |
| Comet Opik | trace/eval/optimization platform |
| Composio | integration fabric |
| Consarc — Noa close agents | finance close |
| CopilotKit Frontend SDK | agent-app state and HITL |
| Coval | voice/conversation simulation |
| Crawl4AI — OSS crawler / Cloud beta | web crawling |
| Credo AI — AI Governance platform | governance evidence |
| CrewAI Crews / Flows | multi agent framework |
| CrewAI — Commercial platform | developer platform |
| Crossmint Agents | wallet card checkout |
| Cua drivers / cloud | desktop controller |
| Cursor — Cloud Agents | software engineering |
| Daily / community — Pipecat | realtime multimodal orchestration |
| Dapr Agents | distributed agent runtime |
| Datadog | APM/agent observability |
| Daytona | executable environment |
| DBOS Transact | durable execution |
| Debezium — PostgreSQL CDC | enterprise data cdc |
| Decagon — AI concierge | customer service |
| Deep Agents | agent harness |
| DeepEval / Confident AI | evaluation harness/platform |
| Deepgram — Voice Agent API | integrated speech agent API |
| DeepSeek-V4-Pro | open weight model |
| Descope — Agentic Identity Hub | delegated authorization |
| Devin | software engineering |
| Dify | visual agent platform |
| Docling library | document parsing |
| Droidrun / Mobilerun — Mobilerun | mobile controller |
| Dropzone AI — AI SOC analyst | secops |
| DSPy | optimization framework |
| Dust | enterprise knowledge builder |
| E2B | executable environment |
| Elastic — Enterprise Search / App Search / Workplace Search | enterprise search |
| Elasticsearch | retrieval substrate |
| ElevenLabs — ElevenAgents | managed voice agent platform |
| Elicit — Research application | research knowledge |
| Evidently | evaluation/monitoring |
| Exa — Search + Contents | web search |
| F5 — AI Guardrails / AI Red Team (CalypsoAI) | runtime ai security |
| FastMCP | MCP framework |
| Finacode — Finance agents / embedded API | finance reconciliation |
| Firecrawl — Cloud / OSS core | web crawling |
| Fireworks AI — Inference / Deployments | managed inference |
| FloQast — Accounting workflows | finance close |
| Flowise | visual agent platform |
| Fly.io Sprites | persistent computer |
| Forge Space — mcp-gateway | tool control plane |
| Fortinet — Virtue AI acquired portfolio | runtime ai security |
| Freeplay | team evaluation workflow |
| Future AGI | integrated quality platform |
| Galileo | evaluation/monitoring |
| Gemini Enterprise Agent Platform | hyperscaler |
| ggml-org — llama.cpp | local edge inference |
| Giskard | evaluation/red-team |
| GitHub — Copilot cloud agent | software engineering |
| Glean — Deep Research | research knowledge |
| Google Agent Development Kit | agent sdk |
| Google Cloud — API Gateway MCP | tool control plane |
| Google — Gemini API / Gemini 3.8 Flash | frontier model api |
| Google — Gemma | open weight model |
| Graphiti | temporal memory |
| Grok API | frontier model api |
| GroqCloud | specialized inference |
| Gumloop | business agent builder |
| Hamming | voice quality tooling |
| Harvey — Workflow agents | legal |
| Hatchet | task workflow execution |
| Heddle | embedded agent runtime |
| Helicone | LLM request observability |
| Heym | specialist self host platform |
| HighRadius — Cash Application / Collections | finance ar |
| Hippocratic AI — Healthcare agents | healthcare |
| Holistic AI — Governance platform | governance evidence |
| HoneyHive | agent quality platform |
| Hugging Face — PEFT | adaptation library |
| Hugging Face — smolagents | code agent sdk |
| Hugging Face — TRL | adaptation library |
| HumanSignal — Label Studio | human annotation |
| Hume — EVI | expressive speech-to-speech |
| Hyperspell — Company knowledge layer | enterprise context connectors |
| IBM ContextForge | tool control plane |
| IBM — watsonx Orchestrate | enterprise builder |
| incident.io — Investigations | it sre |
| IncidentFox — AI SRE | it sre |
| Inferless | managed inference |
| Inngest | durable execution |
| Inspect AI | agent evaluation harness |
| Intercom — Fin | customer service |
| Ironclad — Rivet | visual workflow builder |
| JamJet | action control execution |
| Jina AI / Elastic — Jina embeddings / rerankers | embeddings rerankers |
| Jina AI / Elastic — Reader / Search / Reranker | web acquisition and reranking |
| JSON Schema | interface contract |
| Kong — AI Gateway MCP servers | tool control plane |
| Laminar | agent debugging/replay |
| Langflow | visual developer builder |
| Langfuse | trace/eval platform |
| LangGraph | graph runtime |
| LangSmith Deployment | developer platform |
| LangSmith Observability & Evaluation | agent engineering platform |
| LangWatch + Scenario | simulation + quality platform |
| LastMile AI — mcp-agent | mcp agent framework |
| Letta current agent platform | stateful agent runtime |
| Lindy | business agent builder |
| LiveKit Agents | realtime voice framework and transport |
| LlamaIndex — LlamaParse | document parsing |
| LlamaIndex — Workflows | event workflow framework |
| LLM Guard | runtime ai security |
| llm-d project — llm-d | distributed inference control |
| LMSYS — SGLang | inference engine |
| Locus Pro / Pay With Locus | paid tool access |
| Lunnoa | specialist self host platform |
| Machine Payments Protocol | interface contract |
| Make AI Agents | low code automation |
| Marker | document parsing |
| Mastercard — Agent Pay | card network delegation |
| Mastra | typescript agent framework |
| Maxim | simulation/evaluation |
| Maxim AI — Bifrost | model gateway |
| MCP | interface contract |
| MCP Apps | interface contract |
| MCP Apps ext-apps SDK | embedded app SDK |
| MCP conformance test framework | qualification tooling |
| MCP Tasks | interface contract |
| MCP TypeScript SDK | MCP implementation |
| MCP-UI | UI interoperability library |
| Mem0 — OSS memory / Platform | agent memory |
| Merge — Unified API offering | integration fabric |
| Meta — Llama 4 Maverick | open weight model |
| Microsoft Foundry Agent Service | hyperscaler |
| Microsoft GraphRAG research repository | graph retrieval |
| Microsoft — Agent Framework | agent sdk workflows |
| Microsoft — AutoGen | multi agent framework |
| Microsoft — Azure API Management AI Gateway tier | tool control plane |
| Microsoft — Copilot Studio | enterprise builder |
| Microsoft — Foundry hosted agents | managed compute runtime |
| Milvus / Zilliz — Milvus | retrieval substrate |
| Mistral AI — Mistral Small 4 / Medium 3.5 | open weight model |
| Mistral AI — OCR 4.1 / Voxtral | multimodal specialists |
| Mixedbread — Search/retrieval platform | retrieval ranking |
| MLflow | experiment/lineage platform |
| MLX contributors — MLX LM | local edge inference |
| Modal — GPU Functions / Endpoints | gpu compute |
| Modal — Sandboxes | executable environment |
| ModelOp — AI system inventory | governance evidence |
| Modern Treasury — AI Agent / Workspace | finance payment ops |
| Moonshot AI — Kimi-K3 | open weight model |
| Moveworks | it service |
| n8n | low code automation |
| Nango | integration fabric |
| Nekuda — AgentLane / WebMCP Kit | web action enablement |
| Nevermined Payments | metering and monetization |
| Noma Security — AI Security platform | ai posture runtime |
| Novita AI — Inference / GPU offerings | managed inference |
| Numeric — Close / Agentic Subledgers | finance close |
| NVIDIA — Dynamo | distributed inference control |
| NVIDIA — NeMo Curator | training eval data curation |
| NVIDIA — NeMo Guardrails library | runtime ai security |
| NVIDIA — TensorRT-LLM | inference engine |
| OAuth | interface contract |
| Obot — MCP Gateway | tool control plane |
| OIDC | interface contract |
| Olyra — AP agents | finance ap |
| OPA — Open Policy Agent | policy decision |
| OpenAI Agent Builder | visual agent platform |
| OpenAI Agents API | managed harness runtime |
| OpenAI Agents SDK | agent sdk |
| OpenAI ChatKit | embedded chat ui |
| OpenAI Codex | software engineering |
| OpenAI — ChatGPT Deep research / Work | research knowledge |
| OpenAI — Responses API / GPT-6 Sol, Astra, Luna | frontier model api |
| OpenAPI | interface contract |
| OpenBao | secrets |
| OpenDataLab — MinerU | document parsing |
| OpenFGA | relationship authorization |
| Openfort — Agentic payments | programmable wallet |
| OpenHands — Agent Canvas | software engineering |
| OpenHarness | agent harness |
| OpenLIT | OTel observability |
| OpenRouter — Provider Routing | model access broker |
| OpenSandbox | executable environment |
| OpenTelemetry GenAI | interface contract |
| OpenUI | generative UI language and renderer |
| Oracle — Fusion AI Agent Studio | enterprise builder |
| Oso for Agents | agent authorization |
| PaddlePaddle — PaddleOCR | document parsing |
| Palo Alto Networks — Protect AI acquired portfolio | integrated ai security |
| Paragon — Connect / ActionKit offering | integration fabric |
| Parallel — Search API | web search |
| PayAI — x402 Facilitator | payment verification and settlement |
| Payman — Agentic banking | applied banking agent |
| PayPal — Agent Ready / Store Sync | merchant acceptance |
| Permit.io — Permit platform / MCP Gateway | policy control |
| Pinecone — Hybrid search | retrieval substrate |
| Pipedream Connect | integration fabric |
| PolyAI — Voice agents | contact center |
| Portkey — AI Gateway | model gateway |
| PostgreSQL + pgvector | retrieval substrate |
| Prefect | workflow orchestration |
| Presidio | pii detection |
| Prime Intellect — Verifiers | rl environments |
| Privy Wallet Infrastructure | credential custody and signing |
| Promptfoo | CI eval/red-team harness |
| PromptLayer | prompt lifecycle |
| Prophet Security — AI SOC Analyst | secops |
| Protect AI — ModelScan | supply chain |
| Pydantic AI | agent sdk |
| Pydantic Logfire | application/agent observability |
| Qdrant OSS / Cloud | retrieval substrate |
| Qwen3 Embedding / VL retrieval family | embeddings rerankers |
| Qwen3.8-27B | open weight model |
| Ragas | retrieval/evaluation toolkit |
| Ragie — Managed ingestion/retrieval | enterprise context connectors |
| Reducto — Parse / Extract | document extraction |
| Relevance AI — Agents / Workforces | specialist enterprise builder |
| Restate | durable execution |
| Retell AI | managed telephone agent operations |
| Runlayer — MCP Gateway | tool security gateway |
| Runloop — Devboxes | executable environment |
| Runpod — Pods / Serverless | gpu compute |
| Rye — Universal Checkout API | checkout actuation |
| Salesforce — Agentforce | enterprise builder |
| Sandbox0 | persistent computer |
| SAP — Joule Studio | enterprise builder |
| SentinelOne — Prompt Security | runtime ai security |
| ServiceNow — AI Agent Studio / Fabric | enterprise builder |
| Shopify agentic commerce APIs | commerce state |
| Shopware / Agentic Commerce Alliance — AXP | commerce experience extension |
| Sierra — Customer agents | customer service |
| Sigstore | supply chain |
| Sim | visual agent platform |
| Skills over MCP | interface contract |
| Skyfire KYAPay | identity linked payments |
| Skyvern | managed browser agent |
| SPIFFE — SPIRE | workload identity |
| StackAI | enterprise builder |
| Stagehand | browser agent sdk |
| Steel Browser | cloud browser |
| Strands Agents — Strands harness | agent harness |
| Strands Agents — Strands Harness SDK | agent sdk |
| Stripe Agentic Commerce + Machine Payments | merchant acceptance |
| SwarmLab | specialist self host platform |
| Tambo | controlled generative UI |
| Tavily Web Access APIs | web search |
| Telnyx — AI Assistants | carrier-integrated voice platform |
| Temporal | durable execution |
| TEN maintainers — TEN framework | realtime voice framework |
| TensorZero | model gateway |
| Text Generation Inference (TGI) | inference engine |
| Thinking Machines Lab — Tinker | managed adaptation |
| Together AI — Inference / GPU clusters | managed inference |
| ToolHive | tool control plane |
| Traceloop / OpenLLMetry | instrumentation/quality |
| Trigger.dev | background task execution |
| TruLens | evaluation toolkit |
| turbopuffer | retrieval substrate |
| Twilio — ConversationRelay | telephony speech transport |
| UCP | interface contract |
| Ultravox Realtime | speech-to-speech model platform |
| Unsloth | adaptation tooling |
| Unstructured — OSS / managed ingestion | document ingestion |
| UTCP | interface contract |
| Vapi | managed programmable voice |
| Vectorize — Hindsight | agent memory |
| Vercel Labs — json-render | catalog-driven generative UI |
| Vercel — AI SDK | agent sdk interface |
| Vercel — AI SDK UI | frontend streaming |
| verl project — verl | distributed rl |
| Vespa | retrieval ranking |
| Vic.ai — AP automation | finance ap |
| Visa — Intelligent Commerce | card network delegation |
| vLLM project — vLLM | inference engine |
| Voyage AI / MongoDB — Rerank | reranking |
| Voyage AI / MongoDB — rerank-3 / rerank-3-lite | embeddings rerankers |
| Voyage AI / MongoDB — Voyage-4 / rerank-2.5 | embeddings rerankers |
| Voyage AI / MongoDB — voyage-4-nano | embeddings rerankers |
| Weaviate — Hybrid search | retrieval substrate |
| Weights & Biases Weave | experiment/LLM quality |
| Windsurf | software engineering |
| Workato — Agent Studio / Genies | enterprise automation |
| x402 | interface contract |
| Xians ACP | specialist control plane |
| Z.ai — GLM-5.3 | open weight model |
| Zapier Agents | low code automation |
| Zenity — AI Agent Security | ai posture runtime |
| Zep Cloud / BYOC | managed context |
| Zep Community Edition | agent memory |