deer-workflow vs LangGraph
Graph-engineering runtime where TypeScript owns control flow and replaceable coding-agent CLIs — Codex, Claude Code, Pi — do the semantic work inside each node. Pilot for DeerFlow 3.0. — versus — Build stateful, multi-actor LLM apps as graphs — durable execution, human-in-the-loop, streaming.
Both make the graph rather than the prompt the unit of control. LangGraph is a library whose nodes call models directly, with checkpointing and durable execution; in deer-workflow a node shells out to an entire coding-agent CLI session, so the runtime is orchestrating harnesses, not model calls.
| deer-workflow | LangGraph | |
|---|---|---|
| Stars | 487 | 40k |
| Forks | 52 | 6.8k |
| Language | TypeScript | Python |
| License | MIT | MIT |
| Last activity | 15 days ago | yesterday |
| Topics | orchestration, coding | agents, orchestration |
| Curated connections | 3 | 24 |
deer-workflow — the curator's take
Pick it when the orchestration itself has to be reviewable: phases, inputs, retries and failure handling live in TypeScript you diff in a PR, not in a prompt that hopes the agent remembers the plan. The generator is the hook — `deer-workflow create "<description>"` has Codex apply the bundled workflow-creator skill and prints a runnable module you then edit as normal code, and `--print` streams one JSON event per line so CI can consume a run. Caveats worth knowing before adopting: Bun-only, requires an installed and signed-in agent CLI, ships two examples, and is explicitly a pilot for ByteDance's DeerFlow 3.0 — expect the API to move.
LangGraph — the curator's take
You reach for LangGraph the moment a simple agent loop stops being enough — when you need state that survives a crash, a human approving a step mid-run, or a flow that can loop back on itself. Most teams arrive here from plain LangChain and don't leave. If all you want is a quick tool-calling agent, this is more machinery than you need — start lighter and come back when you hit the wall.