Skip to content

Version Policy

AgentBridge adapters must document the framework versions they target. Agent frameworks move quickly, so every adapter has two version concepts:

  • Adopted baseline: the minimum version AgentBridge intentionally targets in pyproject.toml.
  • Verified version: the exact version exercised in local or CI tests.

Current Adopted Baselines

Package Purpose Adopted Range Verified Locally Status
pydantic Core SDK models >=2.13,<3 2.13.5 Verified
litellm Model-provider delegation >=1.100,<2 1.100.1 Verified install
langgraph Stateful graph backend >=1.2.11,<2 1.2.11 Verified graph + tool execution
pydantic-ai-slim Typed-agent backend import package >=2.42,<3 2.42.0 Verified with offline TestModel
agentbridge-crewai External CrewAI adapter plugin upstream crewai>=0.11.2,<0.12 Scaffolded, not installed Blocked on Python 3.14 dependency resolution
openai-agents External OpenAI Agents SDK adapter plugin >=0.20,<0.21 Latest observed 0.22.2; compatible baseline 0.20.0 Partial native adapter
google-adk External Google ADK adapter plugin >=2.9,<3 2.9.0 Partial native adapter
strands-agents External Strands Agents adapter plugin >=1.55,<2 1.55.1 Partial native adapter
langchain External direct LangChain adapter plugin >=1.4,<2 1.4.0 Partial native adapter
deepagents External LangChain Deep Agents adapter plugin >=0.7,<1 0.7.19 API verified Partial native adapter
agentbridge-agentcore External Amazon Bedrock AgentCore bridge >=0.1,<1 0.1.0 contract verified Runtime and native-service bridge

Notes

  • langgraph is the first real framework adapter verified locally because the current adapter can execute without model credentials.
  • pydantic-ai-slim is preferred over the full pydantic-ai meta-package because AgentBridge adapters should not install every provider/eval/MCP extra just to expose the core pydantic_ai module.
  • CrewAI 0.11.2 failed dependency resolution in this Python 3.14 environment because it depends on older LangChain/LangSmith ranges. Keep the adapter as a skeleton until we either test it in an older Python environment or adopt a newer CrewAI package from an index that resolves cleanly.
  • OpenAI Agents, Google ADK, Strands, and direct LangChain support are intentionally external plugin targets. Their native packages are not part of the core agentbridge install path.
  • OpenAI Agents 0.22.x requires openai>=3, while the current AgentBridge core dependency path through LiteLLM uses openai<3. The OpenAI Agents plugin therefore adopts 0.20.x for executable work first and treats 0.22.x as blocked until the major-version conflict is resolved.
  • OpenAI Agents 0.20.x support validates structured output through native SDK output_type/typed final_output, forwards native Agent options such as handoffs, MCP servers/config, prompts, model settings, hooks, tool-use behavior, and input/output guardrails, plus Runner options such as context, max turns, run config, previous response IDs, sessions, and conversation IDs. Approval interruptions, app-owned approval stores, backend-neutral ApprovalQueue resume payload helpers, resumable-state hints, handoff items, and guardrail diagnostics are normalized into result metadata/events; approval UIs and native SDK resume execution remain extension-level.
  • Google ADK 2.9.0 installs successfully, but it currently pins older OpenTelemetry packages than Strands. Native Google ADK and Strands plugin verification may need isolated environments until their telemetry dependency ranges converge.
  • Google ADK support validates structured output through ADK output_schema plus AgentBridge typed validation, forwards native Agent options such as descriptions, global/static instructions, schemas, generation config, mode, transfer controls, planners, code executors, retry/timeout settings, lifecycle callbacks, and sub-agents, plus Runner options such as session/memory/artifact/credential services and plugins. Run diagnostics preserve event history counts, session/service bindings, transfer-to-agent targets, opt-in session/memory/artifact service snapshots, eval labels/runner bindings, and structured deployment metadata; eval execution and deployment publishing remain extension-level.
  • Strands 1.55.1 support forwards native Agent options such as hooks, plugins, interventions, session/memory managers, context managers, retry strategy, checkpointing, sandbox, storage, and background tasks through StrandsExtension. Native MCP client/tool-provider objects are appended to the Strands tools list and covered by a local MCP stdio server/client fixture; native intervention actions and guardrail traces are fixture-covered while JSON-safe summaries/events are normalized. Plain string MCP labels and deployment targets remain extension metadata until live AgentCore/AWS execution tests are added.
  • LangGraph support validates graph execution, native async ainvoke/astream/resume paths, native batch()/abatch() dispatch, ToolSpec calls, deterministic structured output, conditional routing, checkpoint interrupt reporting, compiled-runtime resume, run diagnostics, and callback/metadata/tag forwarding for LangSmith/Langfuse-compatible runtime instrumentation.
  • LangChain 1.4.0 support validates structured output through native response_format/structured_response, richer stream event normalization, run diagnostics, and forwards native create_agent options such as checkpointers, stores, interrupts, cache, schemas, transformers, middleware, and debug when supplied through LangChainExtension. AgentBridge records memory/retriever hints, includes an offline native retriever/checkpointer/store example, supports LangSmith tracing context and Langfuse's native LangChain callback through the optional observability extra, and does not claim portable memory semantics across frameworks.
  • Langfuse API support follows the current v4/OTLP path: OTLP ingestion sends x-langfuse-ingestion-version: 4, Observations v2 and Scores v3 expose cursor-safe iterators, and legacy ingestion/trace reads remain explicitly compatibility-only.
  • LangSmith Agent Server and deployment support covers assistant introspection, thread/run lifecycle, stateless/background/batch runs, SSE, cron scheduling, deployment revisions, redeploy/interruption, deployment logs, bulk deletion, and tier updates through dependency-free typed transport helpers; hosted execution remains credentialed smoke-test territory.
  • AgentCore 0.1.0 is an external bridge plugin. It normalizes invoke_agent_runtime, preserves Runtime session/trace data, and exposes a generic native service operation escape hatch for Memory, Gateway, Identity, observability, and future AgentCore APIs. AWS deployment and IAM remain control-plane concerns.
  • Latest observed versions for planned plugins were checked with pip index versions on 2026-09-11. They are not verified adapter execution versions yet.
  • If upstream APIs break within a range, tighten the range and update this document in the same change.
  • Never claim full adapter support for a framework version unless a contract test runs against that exact version.

Hosted API Contracts

Hosted API verification is separate from SDK package version verification. The Langfuse pagination baseline was checked against the published cloud OpenAPI schema on 2026-10-06 and is recorded in docs/upstream/langfuse-pagination-contracts.json. It covers ten list endpoints, including native numbered pagination for prompts/datasets/items and cursor pagination for observations/scores/experiments/items. Score-config, annotation-queue, and queue-item pagination are included too. Eight adopted score-config and annotation-queue write contracts are tracked for request, parameter, and successful response shape changes. The weekly API drift check tracks these contracts independently of installed Langfuse SDK versions. Offline tests and the live schema comparison do not prove hosted execution.

Future Framework Matrix

Before adding a new framework adapter, document:

  • Package name and install extra.
  • Adopted version range.
  • Verified version.
  • Core features supported through AgentSpec.
  • Features exposed through capability metadata.
  • Native-only features that require raw backend access.

CLI Verification

Use this command to inspect adopted and installed dependency versions:

agentbridge versions --json