Every automation should leave behind a session you can inspect
Triggers are only the start. The useful record is the request, reasoning, tool calls, result, cost, and follow-up in one place.
By Hakaree Jackson
The short version
- Treat every automated run as an inspectable session.
- Pass secrets to programs without printing their values.
- End idle compute quickly and preserve the useful output.
- Measure retries and tool failures beside model cost.
Make runtime behaviour visible
Agent infrastructure gets expensive in the gaps: idle sandboxes, repeated tool calls, oversized context, and retries that never change the plan. A completed answer hides those details unless the run keeps them attached.
Store the request, tool activity, duration, model cost, and final state as one session. Engineers can then fix the part that failed instead of replacing the whole agent.
Protect the runtime boundaries
Programs may need credentials, but the agent doesn't need to read their values. Provide variable names, pass values through the sandbox environment, and block commands that expose the environment in logs or messages.
Continue reading
Let agents use secrets without teaching them to reveal secrets