Deterministic orchestrator for AI agent workflows. Orchestration is plain Python; agents are leaves. Every non-deterministic event is logged to an append-only audit trail that doubles as a replay tape — giving you resume, rewind, and hot-patch for free.
Pre-stable software. The API may change between minor versions without deprecation warnings. Pin your dependency to a specific version if you use Godel in production.
Most agent frameworks bury orchestration inside the LLM. The agent decides what step comes next by re-reading a plan, occasionally getting it wrong. State survives by luck. When something crashes three agents deep, you start over.
Agent workflows have two parts: structural decisions (what to do next, when to branch, what state to track) and operational decisions (how to implement a fix, whether a review is valid). The first is deterministic and doesn't need an LLM. The second requires judgment and can't be reduced to code.
Godel draws the line. @workflow functions handle structure. Agents handle judgment. Because the orchestrator is deterministic Python, the runtime can event-source every call — and three primitives fall out:
- Resume — a crashed run picks up from the last durable event. Expensive agent calls are not re-paid.
- Rewind — invalidate any event and replay forward. Back up without discarding prior work.
- Hot-patch — edit the workflow, then
godel resume. New code takes effect on the uncached tail.
No other agent framework ships all three.
pip install godel-pyfrom godel import workflow, step, run
@step
def review(pr_number):
return run(["claude", "-p", f"Review PR #{pr_number} for correctness bugs"])
@step
def apply_fixes(review_output):
return run(["claude", "-p", f"Apply these fixes: {review_output}"])
@workflow
def pr_pipeline(pr_number):
findings = review(pr_number)
if "no issues" not in findings.lower():
apply_fixes(findings)godel run pr_pipeline.py -- --pr-number 42Crash at any point. godel resume picks up where it left off — no re-running the review.
| Godel | LangGraph / CrewAI | Temporal / DBOS | |
|---|---|---|---|
| Orchestration | Plain Python | Pre-declared graphs / roles | Workflow/Activity split |
| Resume from crash | Built-in (event log) | Build it yourself | Built-in |
| Rewind to any point | Built-in | Not a concept | Not available |
| Hot-patch live run | Built-in | Redeploy | Redeploy |
| Human-in-the-loop | godel pause / godel repair |
Custom | Custom |
| Agent-first primitives | Session state, schema outputs | Varies | Not designed for agents |
@workflow function # you write this — plain Python
-> @step calls # each step is an event boundary
-> run() / agent() # shells out to Claude, Copilot, or any CLI
-> event log # append-only, deterministic replay backbone
-> resume / rewind / repair
- Why Godel — the thesis in depth
- Getting Started — install, write a workflow, run it
- Concepts — workflows, steps, agents, the audit log, replay
- API Reference — every public symbol
- CLI Reference — the
godelcommand - Examples — annotated walkthroughs
pip install -e ".[dev]"
pytest # 1276 tests across 98 filesPython 3.12+. Conventional commits. CI runs on every push to main.
Business Source License 1.1 — see LICENSE. Use Godel freely as a library in your applications, including commercially. The only restriction is offering Godel itself as a competing hosted orchestration service. Each version converts to Apache 2.0 after three years.