Back to libraryContentsContents

Book

AI Engineering Landscape — September 2026

A dated map of engineering categories, architectural choices and product responsibilities.

About 30 min read

Contents

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

# Category
1 Foundation Models
2 Inference & Model Serving
3 Model Gateways & Routing
4 Agent Frameworks & Harnesses
5 Managed Agent Runtimes
6 Low-code & Enterprise Agent Platforms
7 Agent Tools, APIs & MCP Integration
8 Agent Compute & Sandboxes
9 Browser & Computer Use
10 Search, Web Retrieval & Scraping
11 Document Intelligence & RAG Ingestion
12 Vector & Knowledge Infrastructure
13 Agent Memory & Context
14 Voice & Realtime Agents
15 Evals, Tracing & Observability
16 Runtime AI Security
17 Governance & AI GRC
18 Identity & Agent Authorization
19 Agentic Payments & Commerce
20 Agent Protocols & Interoperability
21 Durable Execution & Business Workflows
22 Human Interfaces, Generative UI & Approvals
23 Model Adaptation & Data Preparation
24 Enterprise Context Sync & Permissions
25 Applied Agent Systems

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