| name | agently-streaming-and-react |
| description | Streaming outputs, runtime streams, and ReAct-style tool loops. |
Agently Streaming and ReAct Skill
Use this skill for streaming outputs, runtime stream UX, and ReAct-style tool loops.
Key Patterns
- Use
get_generator(type="instant") for structured streaming.
- Use TriggerFlow runtime_stream for live UI output.
- ReAct loop: thought -> action -> observation -> final.
Pitfalls to Avoid (Lessons from NexusTodo)
- Do not stop the loop after a
list_tasks if it is only used to locate targets for update/delete; continue to the write action.
- Stop after successful write or query completion to avoid infinite loops.
- Emit task cards only on final/done; stream only thought/observation during execution.
- Keep action enum strict and JSON-only output to prevent parse drift.
- Set LLM
temperature=0 to stabilize structured output.
References
examples/streaming_with_instant_mode.py
examples/response_event_streams.py
examples/trigger_flow_with_agent.py
examples/react_tool_loop.py
Examples
See examples/run.sh for runnable commands.