What ‘state’ should include when an AI agent restarts after failure

For an AI agent, state can mean much more than . When an agent uses tools or changes outside systems, state may include the current plan, tool inputs, , external actions already completed, , policy context, approvals, human edits, handoff notes, and retry or replay choices. If a run stops halfway, logs may not be enough.

The system needs to know what can be safely run again, what needs a fix-up action, and what needs . This makes the design choice important: state could be handled as an app-level object, event log, , trace, or another structure.

Key points

  • AI agent state can include plans, tool results, approvals, human edits, and completed outside actions.
  • A simple log may not show what is safe to run again after a crash.
  • The system needs to separate safe replay from fix-up work or .
  • Possible designs include an app-level object, event log, , or trace.
  • Structured state can help recovery and reduce token use.
Read original