How Agents Work
Follow the lifecycle of an agent from trigger to durable execution, side effects, and final outcome.
Iron Gorilla agents run as governed workflows, not as one giant prompt. A trigger starts work, the agent executes durable steps, tools and data access are mediated by the platform, and each run ends in a clear outcome or escalation.
Lifecycle
- A trigger starts the run.
- The platform loads the agent definition.
- The run progresses through durable steps in order.
- Each step reads input, reasons, calls approved tools if needed, and returns structured output.
- Side effects happen intentionally at explicit boundaries.
- The run finishes with a result, a deferred follow-up, or a handoff for review.
Why durable steps matter
Durable steps are the unit of replay and recovery. If a run is retried after interruption, the platform can resume safely without treating the whole execution as one opaque transaction.
This is why good agent design makes step boundaries explicit:
- one step gathers context,
- another classifies or decides,
- another drafts or prepares,
- and a final step performs or requests an action.
What the platform governs
Iron Gorilla governs the parts of agent execution that create operational risk:
- which triggers are allowed,
- which tools and data scopes are available,
- where side effects happen,
- and how runs recover when work is retried.
The SDK is how you declare that intent in code.