feat(engine): add RemoteAgentFlowEngine for remote agent runtimes

[luyuzhe111]

Summary

This PR adds the integration of AWS Bedrock AgentCore Runtime as a sandboxed remote agent runtime backend where agents and environments collocate. AgentCore Runtime handles isolation and auto-scaling to allow secure and parallel rollouts.

Type of change

• Feature
• Fix
• Docs
• Refactor
• Example / Project
• Infra / CI

What changed

• Introduce RemoteAgentFlowEngine and RemoteAgentRuntime protocol for "agent-in-sandbox" runtimes where agent and environment colocate in a remote container.
• Include AgentCore Runtime adapter using ART's RolloutClient, auto-detection of routable gateway IP for cross-host access, and shared compute_step_metrics utility.
• Gateway now binds to 0.0.0.0 while advertising the routable IP to remote agents to allow access
• Add per-session sampling parameters support in gateway server (different sampling params for train vs val). GatewayManager registers the parameters and selects the right one based on the is_validation flag.
• In pyproject.toml: add agentcore as an extra dependency, add rllm-model-gateway as a uv source to install in editable mode, and enable ruff line check (default is 88 chars, which results in numerous cosmetic wraps; changed to 120 to align with modern convention).

Validation

• pre-commit run --all-files
• Targeted tests: pytest ...
• Manual validation performed
• Not run (reason below)

It seems there are a bunch of pre-commit errors from existing code. might need a separate PR to clean them up.

Validation details:

• uv run pytest tests/engine/test_remote_runtime.py -v passed.
• uv run pytest tests/integration/ -v -s passed locally with AGENTCORE_AGENT_ARN, AGENTCORE_S3_BUCKET, AGENTCORE_BASE_URL, and AGENTCORE_MODEL_ID set.
• e2e training with bash examples/agentcore_math/train_agentcore_math_tinker.sh. The reward curves from rllm-ui are attached below:

Train:

Val:

Docs / examples

• Not needed
• Updated docs
• Updated examples
• Follow-up docs needed