deer-flow
Facts
- Openness
- Open source
- Licence
- MIT
- Language
- Python
- Pinned version
345f08b(2026-09-24)- Topics checked
- 31 of 33
- Kind of agent
- Deep research
What it does notably well
- Compaction
Loaded skill bodies can compact away while paired-read metadata updates a durable skill_context channel that re-injects name/path/description reminders and asks for a fresh read…
- Concurrency
DeerFlow offloads blocking work with asyncio.to_thread, guards against event-loop blocking IO with a Blockbuster-based strict test suite (backend/tests/blocking_io, CI workflow…
- Eval harness
DeerFlow has a record/replay golden: ReplayChatModel returns recorded model outputs matched by caller and input hash through the real Gateway with no API key, and…
- Extension hooks
For an installed enabled plugin contributing observer middleware, isolation tracks downstream invocation and captures its result so an observer failure after execution returns…
- Fresh look each step
DeerFlow keeps the static agent prompt separate and injects date (hidden SystemMessage) and memory (hidden HumanMessage) once, at the first run admission whose history holds no…
- Human approval
When clarification is emitted beside other tool calls, DeerFlow rewrites the AI message to retain only clarification, then returns a stable call-bound question ToolMessage and…
See how deer-flow does on each topic
The topic-by-topic verdicts, the code lines and a side-by-side comparison are in the app. Sign in with Google; new accounts are approved by hand.
Independent analysis, not affiliated with deer-flow.