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(sessionid, action, params, message, image, resumesession)
• Output: response.responsetype + response.responsedata (+ optional response.response_message)
• State slices: request, response, _internal + domain slices (e.g., ticket, orders, workflow)

Recommended Workflow

1. Start from examples/05backendruntime.py.
2. Define your domain slices and register them: NodeRegistry.addvalidslice("your_slice").
3. Implement nodes with NodeContract (keep reads/writes minimal).
4. Build graph with buildgraphfrom_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