---

name: agent-contracts-backend-runtime description: Build an API-oriented agent using AgentRuntime/StreamingRuntime with predictable request/response slices and session support. metadata: short-description: Backend runtime patterns

agent-contracts Backend Runtime

Use this skill when you are implementing an AI agent as a backend service (HTTP API, jobs, or SSE streaming).

Target Shape

• Input: RequestContext(session_id, action, params, message, image, resume_session)
• Output: response.response_type + response.response_data (+ optional response.response_message)
• State slices: request, response, _internal + domain slices (e.g., ticket, orders, workflow)

Recommended Workflow

1. Start from examples/05_backend_runtime.py.
2. Define your domain slices and register them: NodeRegistry.add_valid_slice("your_slice").
3. Implement nodes with NodeContract (keep reads/writes minimal).
4. Build graph with build_graph_from_registry(...) and compile.
5. Wrap with AgentRuntime for request/response execution.
6. If you need progressive updates, use StreamingRuntime and emit SSE via StreamEvent.to_sse().

Guardrails

• Prefer response.response_type for flow termination and client branching.
• Avoid writing to request (discouraged).
• Keep large blobs out of state slices; sanitize before LLM routing (see GenericSupervisor).

References (load only when needed)

• docs/getting_started.md
• docs/core_concepts.md
• docs/cli.md
• docs/skills/official/agent-contracts-backend-runtime/references/patterns.md