Skip to main content

Adapter Compatibility Matrix

FleetForge ships first-party adapters for LangGraph, AutoGen, and CrewAI plus a generic API/OpenAPI contract for bespoke orchestrators. Each adapter runs through the Rust runtime’s agent step so capability tokens, attestations, and deterministic replay behave identically no matter which framework you surface.

Use this page to see what runs today, grab CLI evidence, and follow the upcoming pipeline/serverless/analytics adapters.

LangGraph (canonical)​

Hello Fleet standardises on the LangGraph agent_team DAG and is the only slice we treat as acceptance. /, /demo, and the CLI all reference this run so reviewers always see the same receipts.

  • Runtime invocation: inputs.adapter=langgraph, entrypoint=<module:create_graph>
  • Receipts: Capability tokens mint per step and the adapter signs output whenever TRUST_MESH_ALPHA=1.
  • Replay: Recorded checkpoints replay via ReplayExecutor with byte-for-byte parity.
  • Evidence: cargo test -p fleetforge-runtime golden_path_langgraph_capability_and_attestations (core/runtime/src/lib.rs)

AutoGen (optional)​

Off by default—enable it only when a partner insists on AutoGen parity, then run LangGraph again to reset the storyline.

  • Adapter path: sdk/python/fleetforge/autogen_adapter/
  • Runtime invocation: inputs.adapter=autogen, entrypoint=tests.python.autogen_stub:create_adapter, adapter_inputs={ "task": "..." }
  • Receipts: Scheduler supplies scoped capability tokens and the adapter echoes attestation IDs in summaries.
  • Replay: adapter_smoke_autogen_attestations keeps replays within tolerance.
  • Evidence: cargo test -p fleetforge-runtime adapter_smoke_autogen_attestations (core/runtime/src/lib.rs)

CrewAI (optional)​

Also off by default—only enable the CrewAI adapter when reviewers specifically ask for it.

  • Adapter path: sdk/python/fleetforge/crew_adapter/
  • Runtime invocation: inputs.adapter=crewai, entrypoint=tests.python.crewai_stub:create_adapter, adapter_inputs={ "inputs": {...} }
  • Receipts: Crew runs inherit the same step-scoped capability tokens and all events carry signed provenance.
  • Replay: adapter_smoke_crewai_attestations exercises and replays the sample crew.
  • Evidence: cargo test -p fleetforge-runtime adapter_smoke_crewai_attestations (core/runtime/src/lib.rs)

Custom orchestrators​

Bring your own orchestrator via the Connect/HTTP API and OpenAPI contract:

  • Contract: api/openapi/fleetforge.json
  • Examples: sdk/typescript/examples/semantic-kernel/, sdk/python/fleetforge/agent_adapter.py
  • Receipts: Every custom adapter must emit capability tokens, attestation IDs, and trust.* spans identical to the built-in adapters.
  • Replay: Use fleetforge-ctl submit + fleetforge-ctl replay with the same run_id to verify parity before shipping.

Adapter runner CLI​

Use the shared runner to invoke any adapter entrypoint locally or in CI:

python -m fleetforge.agent_adapter.runner \
  --adapter autogen \
  --entrypoint mypkg.autogen_flows:create_adapter \
  --run-id ${RUN_ID}

Coming soon​

Pipeline/serverless/analytics adapters moved off the homepage—track them here and in the Status & Acceptance promise table until they ship.

  • Pipeline runtimes (Airflow, Argo): Specs align the DAG contract with FleetForge receipts; GA lands after the LangGraph acceptance slice stays green. See the adapter breadth item in the roadmap.
  • Serverless orchestrators (AWS Step Functions): Step Functions adapter shares the same capability/policy envelope; tracked under the adapter breadth roadmap item.
  • Analytics / ELT (dbt): dbt runs inherit trust.* spans + receipts once the analytics adapter clears acceptance; follow the roadmap + promise tracker for status.

New adapters must document their scope here and land smoke tests before we treat them as default.