<!-- Agent Action Record — the no-AI form. Work through the six sections in order
     and keep the record. Read across many runs: patterns in Deviation and Change
     tell you more than any single review. Concept from the Agent After-Action
     Review skill (github.com/quarterback/AAR); shipped here under MIT. -->

# Agent Action Record

**Review date:**
**Reviewed by:**
**Agent / system:**
**Task or period under review:**
**Accountable human (name or role):**

---

## 1. Intent — what was supposed to happen

**Instruction given (actual wording):**

**Success defined in advance as:**
<!-- or: "not defined in advance" -->

**Authority granted to the agent:**
<!-- what it was allowed to decide/act on without a human -->

**Out of scope:**

---

## 2. Action — what actually happened

**What the agent did (step by step, from logs where available):**

**What it produced or changed:**
<!-- files, messages, records, actions in the world -->

**Where actual behavior differed from the instruction:**

**Stayed within granted authority?**  ☐ Yes  ☐ No  ☐ Unknown
<!-- if no/unknown, explain -->

---

## 3. Judgment — human in the loop

**Decisions the agent made on its own:**

**Of those, which should a human have made or seen:**

**Where a human actually intervened (approve / edit / override / redirect):**

**Points where it should have escalated and didn't:**

**Accountable for these actions:**

---

## 4. Deviation — the gaps

**Gaps between intent and action, and why (root cause, not first answer):**

**Good deviations (correct departures from a flawed instruction):**

**Confidently wrong (high confidence + wrong — flag specifically):**

---

## 5. Consequence — what it cost or risked

**Actual outcome (did the work hold up):**

**Harm / cost / risk — realized or narrowly avoided:**

**Downstream affected parties:**

**Expected failure if this run happened 100 times:**

---

## 6. Change — what happens next

**Specific changes before next run (instruction / scope / authority / escalation / checkpoints):**

**Keep doing:**

**Should not be delegated to an agent at all:**

**Signal that the change worked:**