Deep Agents Integration

AgentBridge exposes LangChain Deep Agents through the optional external agentbridge-deepagents plugin. The plugin keeps Deep Agents' filesystem, sandbox, interpreter, and deployment dependencies outside the core install while preserving the native option surface.

Version Policy

Package	Adopted range	API baseline	Boundary
deepagents	>=0.7,<1	0.7.19	External runtime plugin

Install it only when needed:

pip install -e "plugins/agentbridge-deepagents[runtime]"

Portable Definition

from agentbridge import AgentSpec, ToolSpec, run_agent

agent = AgentSpec(
    name="research_agent",
    instructions="Research the request and cite the files you inspect.",
    model="openai/gpt-5",
    tools=[ToolSpec.from_function(search_catalog)],
    backend_config={
        "deepagents": {
            "skills": ["/workspace/skills/research/"],
            "memory": ["/workspace/memory/AGENTS.md"],
            "permissions": [{"mode": "deny", "path": "/workspace/secrets/**"}],
            "backend": filesystem_backend,
            "native_options": {"custom_profile": "research"},
        }
    },
)

result = run_agent(agent, backend="deepagents", input="Find the refund policy.")

Native Surface Mapping

The adapter forwards these create_deep_agent options from agent.backend_config["deepagents"]: middleware, subagents, skills, memory, permissions, backend, checkpointer, store, interrupt_on, response_format, context_schema, and general_purpose_subagent. A native object is not converted or reimplemented; it is passed through unchanged. Other compatible options can be supplied with native_options, with collisions against portable fields rejected explicitly.

Run context is passed through Deep Agents' native context= channel. Session IDs become runtime thread_id configuration, and observability metadata/callbacks use the same AgentBridge configuration helpers as the LangChain and LangGraph adapters.

Boundary And Tests

The plugin has credential-free contract tests for model normalization, tool translation, native option forwarding, context propagation, and normalized streaming. Live provider execution, sandbox security, filesystem persistence, remote deployment, and hosted tracing still require the corresponding credentials or service fixtures. Those are intentionally separate smoke lanes rather than pretending that a local fake proves a hosted integration.

uv run pytest -q plugins/agentbridge-deepagents/tests
uv run ruff check plugins/agentbridge-deepagents/agentbridge_deepagents plugins/agentbridge-deepagents/tests