Steps and Replay
Model agent work as durable steps so retries, recovery, and side effects stay predictable.
Durable steps are how Iron Gorilla turns agent logic into recoverable production workflows. They are the boundary between “reason about work” and “perform work.”
How to think about step boundaries
A good step does one coherent unit of work and returns structured output that the next step can use.
Typical flow:
- Gather context.
- Decide or classify.
- Draft or prepare an action.
- Perform the side effect or request approval.
Why this matters
When steps are explicit:
- retries are safer,
- failures are easier to inspect,
- side effects are easier to control,
- and the execution model stays understandable.
What to avoid
- Long steps that both reason and perform many actions.
- Hidden writes buried inside what looks like a read-only phase.
- Steps that depend on implicit mutable state instead of explicit inputs and outputs.