Conformance

AgentBridge conformance checks verify the minimum contract every adapter must satisfy before it can claim baseline compatibility.

Run all discovered backends:

agentbridge conformance

Run one backend:

agentbridge conformance --backend langchain
agentbridge conformance --backend google_adk --json

What It Checks

• Capability metadata exists and names the selected backend.
• A simple no-tool agent compiles and returns normalized output.
• Streaming returns normalized events and ends with complete.
• Sync tools work when the adapter advertises tools.sync: full.
• Structured output works when the adapter advertises structured_output: full.
• Run diagnostics metadata exists when the adapter advertises observability.diagnostics: full.

Passing conformance means the adapter satisfies the portable AgentBridge contract. It does not mean every native framework feature is fully abstracted.

Offline Plugin Models

Cloud-backed frameworks often require provider packages, API keys, or cloud credentials. Their AgentBridge plugins therefore provide a plugin-only agentbridge/offline model string for conformance. This string is for tests and contract checks only.

Current offline conformance paths:

Backend	Native path still exercised	Offline model implementation
openai_agents	OpenAI Agents SDK Agent and Runner	Local SDK Model subclass
strands	Strands Agent event loop	Local Strands Model subclass
langchain	LangChain create_agent graph	Local LangChain BaseChatModel subclass
google_adk	Google ADK Agent and Runner	Local ADK BaseLlm subclass

Current Snapshot

As of the latest local sweep, agentbridge conformance passes for every discovered backend.

Backend	Baseline	Tools	Streaming	Structured Output	Diagnostics
google_adk	pass	pass	pass	pass	pass
langchain	pass	pass	pass	pass	pass
langgraph	pass	pass	pass	pass	pass
mock	pass	pass	pass	skipped, not advertised as full	skipped, not advertised as full
openai_agents	pass	pass	pass	pass	pass
pydantic_ai	pass	pass	pass	pass	skipped, not advertised as full
strands	pass	pass	pass	pass	pass

Production users should keep using normal framework model identifiers, such as openai/gpt-5, anthropic/claude-sonnet, or framework-native model objects. Adapter-specific offline models are not provider-routing abstractions.

Relationship To Feature Parity

Regular push and pull-request CI runs the LangChain plugin suite on Python 3.11 and 3.12. This includes dependency-free API request contracts, incremental HTTP streaming against a local server, remote interruption/reconnect normalization, and real local LangGraph checkpoint/resume tests. It also executes the offline streaming and approval examples. Credentialed hosted smoke checks remain separately gated.

Conformance is the floor. Feature parity is tracked through capability reports, adapter docs, and framework-specific issues.

Use these commands together:

agentbridge conformance --backend langchain
agentbridge coverage-report --backend langchain --markdown
agentbridge extensions langchain --json

Only upgrade a capability to full when a native-framework test proves the behavior. Keep framework nuance as extension or native_only when forcing it into the common AgentSpec would be misleading.