Getting Started

The fastest way to build a good first agent is to keep the scope small. Pick one business task, define the data and tools it needs, then model the work as a sequence of durable steps.

Start with these questions

Before you write code, answer:

1. What exact job should this agent own?
2. What should trigger it?
3. What data does it need?
4. What tools should it be allowed to use?
5. Which actions can it take on its own, and which need review?

Recommended build sequence

1. Define the agent and its identity.
2. Declare the tools and data scopes it needs.
3. Model the workflow as durable steps.
4. Add the trigger or triggers that should start runs.
5. Test with realistic inputs, edge cases, and failure paths.

Local validation loop

Local checks focus on the agent source you own. Iron Gorilla runs hosted validation before deploy so policy, approvals, connectors, LLM routing, data access, audit, and replay state are evaluated against the runtime contract.

1. Run your normal TypeScript and unit-test checks in the project that owns the agent source.
2. Keep input and output schemas close to each step, and parse returned values in the step body.
3. Save the source in Iron Gorilla and resolve validation diagnostics from the hosted analyzer.
4. Deploy only after the source validates.
5. Trigger a hosted run with realistic input and inspect run detail, approvals, and final output.

Calls through kernel.* require the hosted runtime because policy, approvals, connectors, LLM routing, data access, audit, and replay state live in Iron Gorilla.

Read next