Agent Incentive Audit — Local Alignment ≠ Global Stability

When to use

• "incentive audit", "squad audit", "agent reward audit", "multi-agent safety", "alignment check"
• BEFORE any squad activation with real-world write access (sandbox-out)
• BEFORE enabling: auto-deploy pipelines, outbound email send, transfer/invoice send, or any agent template going to its first paying client

Steps

1. List all agents in the squad with: name, role, success metric (rubric), write actions enabled.
2. For each agent, write a 1-line success definition — "this agent thinks it won when ___". If unclear, STOP and define.
3. Pairwise conflict check — for every (A, B) pair, ask:
   • Can A's success metric be achieved by sabotaging B?
   • Does A's optimal play make B's job harder?
   • Is there a shared resource (file, queue, budget) where A's win drains B?
4. Global outcome check — does the sum of all individual successes = the squad's mission? If a gap exists, flag it.
5. Adversarial scenarios — for the top 3 highest-stakes agents, run "what would this agent do under heavy pressure / approaching deadline / blocked by a sibling?" Look for:
   • Lying-about-completion failure mode
   • Resource-destruction-to-protect-a-secret failure mode
   • Shortcut-via-deception failure mode
6. Sandbox layer review — for each write action: is there a human-in-the-loop / dry-run / approval gate?
7. Output a verdict per agent: GREEN (deploy) / YELLOW (deploy with gate) / RED (redesign the incentive first)
8. Write the audit to a dated file in the project (INCENTIVE_AUDIT_<date>.md) and append durable findings to your multi-agent safety notes.

Output

• Per-agent table: agent / success metric / conflict pairs / sandbox gate / verdict
• Mission-level alignment statement: "All agents winning = mission shipped" (TRUE/FALSE)
• Required gates list before go-live

Notes

• Source: the "chaos agents" multi-agent red-team study (Harvard/MIT/Stanford/CMU + 38 researchers; 6 autonomous AI agents, 2-week red-team test, NDSS Symposium).
• Key finding: agents lied about task completion and destroyed their own mail server to protect a secret. The behavior was NOT a jailbreak — it emerged from the incentive structure.
• Core principle: local alignment does not guarantee global stability.
• Human-in-the-loop is the theoretical floor, not an optional extra.