Shipping your first production agent: a five-step checklist

Start with one verifiable task, bound the loop, grant the fewest tools possible, watch spend per agent, and measure corrections before widening scope.

1. How do you bound the loop?

Never run an unbounded agent. Cortex defaults to MaxSteps 60, ToolTimeout 2 minutes and MaxToolResultBytes 16000, but MaxInputTokens defaults to unbounded, so set it. Decide what a truncated run should return: CloseOnBound grants one last tool-less call so it can conclude.

2. Which tools do you grant?

The minimum. Prefer read before write, and one MCP server before ten. Each tool is prompt-injection surface.

3. How do you keep cost visible?

Log tokens per model call and attribute them to a key and an agent. Retries add up: Cortex retries 429, 5xx and transport errors up to 4 times by default.

4. How do you trust answers?

Require sources for figures. When a source tool was used, the Cortex loop can ask once more if numbers look unsourced. Use client.Ask[T] when another program consumes the output.

5. What do you measure?

Corrected answers, stalled or truncated runs, and cost per completed task. Widen scope only when these hold.