Adapter Target Research

This document tracks next adapter targets beyond the first v0 backends. It should be updated whenever new framework research changes adapter priority, package names, or capability expectations.

Research date: 2026-09-11. Package index check: 2026-09-11.

Summary

Recommended priority:

1. agentbridge-openai-agents
2. agentbridge-strands
3. agentbridge-langchain
4. agentbridge-google-adk

Rationale:

• OpenAI Agents SDK maps closely to AgentBridge's current concepts: agents, tools, handoffs, guardrails, runner/results, tracing, and human approval.
• Strands is important for AWS-native production agents, hooks, MCP, conversation managers, structured output, observability, and AgentCore/Lambda-style deployment paths.
• Direct LangChain support complements the built-in LangGraph adapter for teams with existing LangChain agents, middleware, callbacks, memory, retrievers, and LangSmith-style observability.
• Google ADK is strategically important for enterprise-scale, multi-language, deployment-oriented agent systems and has strong session/memory/deployment concepts.

Target Matrix

Target	Package Target	Distribution	Priority	Why
OpenAI Agents SDK	agentbridge-openai-agents	External plugin	P0	Close conceptual fit with AgentBridge events, tools, handoffs, guardrails, tracing, approvals.
Strands Agents	agentbridge-strands	External plugin	P1	AWS-native production agent path with hooks, MCP, structured output, observability, and AgentCore alignment.
LangChain	agentbridge-langchain	External plugin	P1	Direct compatibility for existing LangChain agent apps beyond the built-in LangGraph adapter.
Google ADK	agentbridge-google-adk	External plugin	P2	Strong enterprise, deployment, session/memory, multi-agent story across Google ecosystem.

Adopted Version Targets

Framework	Native Package	Adopted Range	Latest Observed	Status
OpenAI Agents SDK	openai-agents	>=0.20,<0.21	0.22.2; compatible baseline 0.20.0	Partial native adapter
Strands Agents	strands-agents	>=1.55,<2	1.55.1	Partial native adapter
LangChain	langchain	>=1.4,<2	1.4.0	Partial native adapter
Google ADK	google-adk	>=2.9,<3	2.9.0	Partial native adapter

OpenAI Agents SDK

Sources:

• OpenAI Agents SDK docs
• OpenAI Agents on developers.openai.com
• openai/openai-agents-python

Observed concepts:

• Agent definitions.
• Runner-driven execution.
• Tools.
• Handoffs.
• Guardrails.
• Results and state.
• Integrations and observability.
• Human approvals and durable/session integrations.

AgentBridge mapping:

AgentBridge Area	OpenAI Agents Mapping
AgentSpec.instructions	Agent instructions.
ToolSpec	SDK tools/function tools.
AgentEvent	Runner events, tool calls, handoffs, guardrails, result lifecycle.
structured_output	Agent/output type support where available.
human_approval	Approval/human-in-the-loop flows.
observability.raw	Native traces and run objects.
workflow.delegation	Handoffs.

First plugin shape:

agentbridge scaffold-plugin plugins/agentbridge-openai-agents --backend openai_agents

Initial extension namespace target:

OpenAIAgentsExtension.config(
    handoffs=[...],
    guardrails=[...],
    tracing=True,
    approval_policy={...},
    approval_store=ApprovalQueue(),
)

Acceptance path:

• Compile simple AgentSpec into SDK Agent.
• Map ToolSpec callables to SDK tools.
• Run with the SDK runner and normalize final result.
• Normalize tool-call and completion events.
• Persist approval interruptions into app-owned stores or ApprovalQueue resume payload helpers.
• Preserve raw run/result objects.
• Add conformance tests with mocked/offline model paths if available.

Risks:

• The SDK may evolve quickly.
• Some runtime features may require OpenAI credentials or hosted services.
• Handoffs and guardrails should be extension-level before becoming common API.
• openai-agents 0.22.x requires openai>=3; current AgentBridge core uses LiteLLM with openai<3, so executable adapter work starts on 0.20.x.

Google ADK

Sources:

• ADK official site
• ADK get started
• google/adk-python
• Google Cloud ADK overview
• Google Cloud ADK state and memory article

Observed concepts:

• Code-first agent framework.
• Agents and tools.
• Debug, evaluate, and deploy flows.
• Multi-agent systems.
• Sessions and memory.
• Enterprise deployment and Google ecosystem alignment.
• Multiple language SDKs.

AgentBridge mapping:

AgentBridge Area	Google ADK Mapping
AgentSpec.instructions	Agent instruction/prompt definition.
ToolSpec	ADK tools.
state.session	Sessions.
memory.long_term	ADK memory/storage options.
workflow.delegation	Multi-agent orchestration.
observability.raw	Native ADK run/session objects.
deployment	ADK deployment surfaces.

First plugin shape:

agentbridge scaffold-plugin plugins/agentbridge-google-adk --backend google_adk

Initial extension namespace target:

GoogleADKExtension.config(
    app_name="support",
    session_service="memory",
    memory_service="memory",
    artifact_service="local",
    sub_agents=[...],
    capture_service_snapshots=True,
)

Acceptance path:

• Compile simple AgentSpec into an ADK agent.
• Map sync ToolSpec callables into ADK tool functions.
• Support session id propagation from RunInput.session_id.
• Capture opt-in session/memory/artifact service snapshots through native ADK-shaped service methods.
• Preserve native session/run output as raw.
• Add capability metadata for sessions and memory.

Risks:

• ADK spans multiple languages and deployment targets; Python plugin should stay focused first.
• Session/memory services may require environment-specific setup.
• Google ecosystem deployment features should be extension-level.
• ADK 2.9.0 currently pins older OpenTelemetry packages than Strands, so combined native adapter installs may require isolated environments.

LangChain

Sources:

• LangChain Python docs
• LangChain agents docs
• LangChain middleware docs

Observed concepts:

• Agents built on LangGraph.
• Tool calling and tool strategies.
• Middleware.
• Model/provider abstraction.
• Short-term memory and persistence patterns.
• Retrieval and RAG chains.
• Callbacks, tracing, and LangSmith ecosystem.

AgentBridge mapping:

AgentBridge Area	LangChain Mapping
AgentSpec.instructions	System prompt / agent prompt template.
ToolSpec	LangChain tools.
tools.async	Async tool and runnable support.
state.memory	LangChain memory/checkpointer patterns.
observability.tracing	Callbacks and LangSmith-style tracing.
workflow.graph	Prefer built-in LangGraph backend for graph-native orchestration.

First plugin shape:

agentbridge scaffold-plugin plugins/agentbridge-langchain --backend langchain

Initial extension namespace target:

LangChainExtension.config(
    agent_type="tool_calling",
    middleware=[...],
    callbacks=[...],
    memory="conversation_buffer",
    retrievers=[...],
)

Acceptance path:

• Compile simple AgentSpec into a direct LangChain agent.
• Map ToolSpec callables into LangChain tools.
• Support middleware and callbacks through extension config.
• Preserve raw agent/result/callback metadata.
• Clearly distinguish direct LangChain support from the built-in LangGraph adapter.

Risks:

• LangChain agents are themselves built on LangGraph, so boundaries between langchain and langgraph adapters must stay clear.
• Model provider behavior may overlap with LiteLLM; AgentBridge should not build a second model abstraction.
• Retrieval/memory integrations should stay extension-level until common semantics are proven.

Strands Agents

Sources:

• Strands Agents official site
• Strands structured output docs
• AWS Strands Agents technical deep dive

Observed concepts:

• Python and TypeScript SDKs.
• Agent and tool decorators.
• Conversation managers.
• MCP clients/tools.
• Hooks for tool lifecycle, interventions, and guardrails.
• Structured output.
• Observability and trace attributes.
• AWS/Bedrock/AgentCore deployment alignment.

AgentBridge mapping:

AgentBridge Area	Strands Mapping
AgentSpec.instructions	system_prompt.
ToolSpec	@tool or tool objects.
structured_output	structured_output_model / structured output schema.
streaming.events	Hook/event lifecycle.
human_approval	Hook-based interrupt/approval patterns.
state.session	Conversation managers.
observability.raw	Trace attributes and native agent/run objects.
tools.mcp	MCP client tooling.

First plugin shape:

agentbridge scaffold-plugin plugins/agentbridge-strands --backend strands

Initial extension namespace target:

StrandsExtension.config(
    conversation_manager="sliding_window",
    trace_attributes={"service": "support-agent"},
    hooks=[...],
    mcp_clients=[...],
)

Acceptance path:

• Compile AgentSpec into Strands Agent.
• Map ToolSpec callables into Strands tools.
• Map output_type to structured output path.
• Normalize hook/tool lifecycle events into AgentEvent.
• Preserve native MCP clients, intervention actions, and guardrail traces while exposing JSON-safe summaries.
• Preserve trace attributes and raw result.

Risks:

• Python and TypeScript APIs may diverge.
• AWS deployment integrations should remain extension-level.
• Hook-driven control flow may need a richer AgentBridge event model.

Capability Gaps To Add Before These Adapters

• tools.mcp: added to canonical taxonomy.
• tools.openapi: added to canonical taxonomy.
• guardrails: added to canonical taxonomy.
• workflow.handoffs: added to canonical taxonomy.
• deployment.serverless: added to canonical taxonomy.
• observability.tracing: added to canonical taxonomy.
• observability.diagnostics: added to canonical taxonomy.
• state.memory: added to canonical taxonomy.
• runtime.retries: added to canonical taxonomy.
• evals: added to canonical taxonomy.

Recommended Next Work

1. Add OpenAI Agents native SDK resume execution fixtures around stored approval queue payloads and SDK run state.
2. Add Strands live AgentCore/AWS deployment smoke coverage beyond metadata preservation.
3. Add additional real-world LangChain/LangGraph import fixtures as new project shapes appear beyond the current RunnableSequence and compiled StateGraph coverage.
4. Add Google ADK eval execution hooks and deployment publishing examples beyond metadata preservation.
5. Keep import/migration helpers for existing LangChain and LangGraph apps conservative, expanding only when object shapes can be inspected safely.
6. Use agentbridge conformance --all, coverage reports, extension schemas, and issue checklists as the release gate before claiming broader framework parity.

Current migration helper status:

• import_langchain_agent() extracts obvious name, model, prompt, tools, runnable methods, LCEL steps, graph summaries, schema names, interrupt hints, native storage/checkpoint hints, and extension hints from LangChain-like objects.
• import_langgraph_graph() extracts graph names and visible node names while marking topology as native-only/manual-review territory.