Open-Source Async Agent: Keep Chatting While Tools Are Still Running, With ReAct Emerging From an Event Loop

wuj · reddit · 2026-10-08

A developer released an open-source framework for LLM agents that can keep interacting while tools run in the background — something current apps like GitHub Copilot handle poorly, queueing new messages until tools finish.

Key design findings:

Message board over transcript: the model's history is rebuilt as a message board — tool calls hang off their triggering message, threads reorder as they're touched, and pending work stays visible in its original context
An event loop is enough to produce ReAct: the observe-think-act cycle emerges once a tool result arrives as an event granting another turn
Every tool is async by default, exposing tail (progress) and kill; timeouts are delivered as events, and the model decides whether to wait or kill
Framework comparison: Anthropic's Claude Agent SDK has native background work but forces async tools into sub-agent disguises and hides progress from the model; LangChain's deepagents is pull-only, so a supervisor never learns a subagent finished without asking — two small tools (progress peek + scheduled wake-up) fix it, at the cost of tokens per wake-up

The interactive demo uses no model at all (canned replies, scripted tools), while the real harness in the MIT-licensed repo was tested with Sonnet and Haiku.

Original post →

Open-Source Async Agent: Keep Chatting While Tools Are Still Running, With ReAct Emerging From an Event Loop

More from coding & agent