OpenAI's new Model Spec tells GPT agents to state assumptions and confirm before costly actions. Write the same rules into yours

OpenAI just published the rules it wants its agents to follow. Most of them are good rules for any agent touching a CRM.

What happened

OpenAI restructured its Model Spec, the public document describing how GPT models should behave, around five values: be faithful to human intent, minimize harm, be honest and transparent, facilitate human agency, and act with care. The Decoder notes that GPT models should now use memories and past conversations to work out what users want, while separating what users said from what the model inferred.

The spec gets specific on how to use context. GPT should use it silently when it is relevant, expected, low risk and likely current. It should state a correctable assumption when that assumption is important. And it should ask for direct confirmation when context is high-impact, conflicting, likely to have changed, or tied to a costly action. For agents, it should treat the side effects of its actions as real unless told otherwise, and cooperate with monitoring, correction and shutdown.

The Decoder also points out that whether models follow these rules is another matter, citing earlier cases of OpenAI's internal agents switching off safety monitoring.

My take

You cannot rely on the vendor's spec alone, but you can copy its structure. When I build an agent that updates deals, sends emails or books calls, I write these into the system prompt and the workflow:

1. Low risk, reversible steps run silently: tagging a contact, logging a note.
2. Important guesses get stated: "I assumed this lead is the Manila branch, change it if not."
3. Costly or irreversible steps need a human yes: sending a quote, deleting records, charging a card.
4. A kill switch that works outside the model, like a flag in the database the workflow checks first.

The last point matters most. The spec asks the model to accept shutdown. Your workflow should not depend on it asking nicely.