The review
What it does
Gantry is a developer framework for building agents that handle long, multi-step tasks. Its core idea is simple: the agent keeps its plan, progress and open questions in a plain-text file that both the agent and you can read and edit. Steps can be retried, skipped or handed to a human without losing the thread.
Why we like it
Long tasks fail because small errors compound and agents lose track of where they are. Gantry's plan file attacks both problems, and makes debugging feel like reading a colleague's notes. Documentation is clear, the maintainers are responsive and it works with most major model providers.
Watch out for
- It's a framework, not a product. You'll be writing code.
- The plan file can get long on big tasks; prune it or the agent spends tokens rereading history.
- Observability beyond the plan file means wiring up your own logging.