[ananthsub]

What does this PR do?

Defines named direct-HTTP and streamable-HTTP MCP access independently of any session lifecycle. These contracts carry episode-scoped connectivity and authorization data; locally launched MCP processes remain Agent Server configuration.

Single-agent interaction flow

This diagram shows the built-in single-agent use case. Other Environment Server protocols may coordinate different participants and use different request, result, and lifecycle flows.

sequenceDiagram
    participant C as Rollout collector
    participant E as Environment Server
    participant R as Task / Resources Server
    participant A as Agent Server
    participant M as Model Server
    participant S as Task sandbox

    C->>E: POST /run
    E->>R: Seed task session
    R-->>E: Tool access and optional SandboxAccess
    E->>A: Seed agent session
    A->>S: Connect as borrower when supplied
    E->>A: Invoke /v1/responses
    A->>M: Model and capture requests
    A-->>E: Agent response
    E->>A: Close agent session
    E->>R: Verify response
    E->>R: Close task session
    E-->>C: Episode result

Loading

The Environment Server coordinates ordering and the final result. Agent and Task/Resources Servers retain ownership of the state and resources they create.

This is part 4 of 14 in the Environment Server stack. Previous: #3484. Next: #3545.

Validation

• uv run pytest tests/unit_tests/test_tool_access.py -q

Rollouts

See the final PR for end-to-end rollout evidence.

Compatibility and benchmark impact

Additive contract definitions only. Static stdio MCP configuration remains owned by individual Agent Server implementations.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.