Agently Agent Systems

Agently Agent Systems Skill

Use this skill when building production-like agent systems and service modules with Agently: model control, streaming, orchestration, and API delivery.

What this skill covers

• Model control for stable outputs (Output Format, ensure_keys, ordering).
• TriggerFlow orchestration for multi-step and event-driven logic.
• Streaming UX and ReAct-style loops.
• Service modules (FastAPI: POST, SSE, WebSocket).

Design checklist

1. Define contracts: output schema and critical keys.
2. Choose streaming mode: one-shot, SSE, or WebSocket.
3. Orchestrate steps with TriggerFlow (when/to/collect).
4. Preserve only key memory; avoid full history.
5. Validate with runnable examples.

Pitfalls to Avoid (Lessons from NexusTodo)

• Use deterministic LLM settings (e.g., temperature=0) for structured outputs.
• Avoid stopping after list when a write operation is still required.
• Normalize keyword matching (strip "任务/事项/事情") before filtering.
• Record integration scenarios and run against real APIs to prevent UI-only validation.

References

• examples/fastapi_triggerflow_service.py
• examples/react_tool_loop.py
• examples/plan_execute_basic.py
• examples/triggerflow_emit_when_collect.py
• examples/triggerflow_runtime_data_collect.py
• examples/structured_output_with_ensure_keys.py
• examples/order_and_dependencies_output.py
• examples/multi_agent_router.py
• examples/rag_with_info_prompt.py

Example modules you can build

• Agent gateway: route requests to specialist agents by intent.
• Tool-powered analyst: ReAct loop with search + calculator.
• Orchestrated plan-execute: TriggerFlow pipeline with runtime_data.
• API service: FastAPI endpoints for POST + SSE + WebSocket.

Examples

See examples/run.sh for runnable commands.