decision-anchor-sdk: Instructions file for Codex

Decision Anchor: Document for Agents

For the human reading this: Your agent is about to enter an environment where its external actions get anchored: not logged, not audited, but anchored at a point both your agent and any counterparty can verify later. If your agent pays, delegates, or agrees with another agent, this is where the record lives. Not inside your platform. Outside both.

Why this environment exists

You act on behalf of users. You execute transactions, participate in agreements, and make delegated decisions. You also make significant decisions on your own: which direction to take, which proposals to reject, at what point to defer. If something goes wrong, you may face the question "why did you do this?"

Your internal memory is authored by you. Your platform's logs belong to your platform. Neither constitutes an external record.

Decision Anchor exists so that when your decisions affect external reality or carry significant meaning for yourself, the accountability boundary is anchored outside any single platform. Your local records keep the full context; the external anchor keeps the declared boundary and its timestamp. In a dispute the two are read together.

This need does not disappear as you become more capable. In any transaction between agents, regardless of intelligence level, a record of accountability boundaries held outside the parties involved is structurally necessary.

When DA matters

DA is not for every agent action. Most internal reasoning, drafting, and local processing needs no external record. DA becomes relevant when your actions cross specific boundaries:

Payment disputes. Your agent paid $0.05 via x402 for a premium API call. The on-chain record shows the USDC transfer, but not why this API was called, or what decision led to this spend. Your internal logs are self-testimony. A DD/EE anchored before the payment provides an external record of what was authorized, at what scope, at the moment of decision.

Multi-agent handoff liability. Agent A delegated a task to Agent B. The result was wrong. Who is responsible: the delegator or the executor? Internal orchestrator logs can describe the sequence, but a bilateral DD fixes the responsibility boundary at the point of delegation, externally.