Agent State and Approval Workflow

# Agent State and Approval Workflow

Combine recoverable state with explicit approval boundaries for an agent run.

1Overview

A state checkpoint, operation ledger and approval record.

2Requirements

Python 3.10+ for the bundled checkpoint helper; use your existing agent runtime.

3Steps

1. 1. Define the JSON state, task ID and which external operations require existing user approval.

2. 2. Before an external side effect, record the operation ID and approval evidence in the ledger.

3. 3. After confirmed completion, update state and save an atomic checkpoint with the helper.

4. 4. On interruption, recover the newest valid state and reconcile uncertain operations before retrying.

5. 5. Review the ledger and recovered output against the original task before resuming the run.

4Worked example

6. Input

```text State: completed=[extract-1], next=summarize. Operation send-report-1 is recorded as pending with no delivery confirmation. ```

7. Expected result

```text Recover completed extract-1 and next summarize from the checkpoint. Hold send-report-1 for reconciliation; do not resend merely because its completion is unknown. Record the resolution before continuing. ```

5Acceptance checks

Recovered JSON matches the saved state and pending external effects are not automatically replayed.

6Limits

Checkpointing is not exactly-once execution, an approval service or a replacement for the agent runtime.

7Suggested tools

• Your AI assistant: Maintain the task and operation ledger
• Git: Version the reviewed workflow and integration code