e27 published a piece on 12 November 2026 from a community contributor who works on AI-assisted software delivery. The author describes a hypothetical mobile app team. The team decides people must be able to record a workout without an account. One agent builds local storage. A second adds synchronisation and puts a login screen in front of the workout. A third writes tests that confirm the login works.

Each task looks fine on its own. Together they undo the product decision. The handoff said “local storage complete” and dropped the reason that made the work correct.

The claim is that handoffs need more than a summary

The author’s fix is a short decision record that travels with the work. It states the decision, the rationale, its status, who may change it and where the supporting material sits. Handoffs separate what is done, what is decided and what is still open. Acceptance criteria come from the decision before anyone builds. One line from the piece says it well: “A summary can point to a decision” and it “should not quietly become the only surviving evidence of one.”

I think the diagnosis is right, and I’d ask the same question I ask of any AI story. Is there a real need? Yes. Anyone who has handed work from one agent to another, or from one person to another, has seen a constraint go missing along the way.

What changed and what didn’t

What changed is the speed. Agents finish tasks fast and hand over code that runs. A missing decision used to surface in a meeting. Now it surfaces after three agents have built on top of it.

What didn’t change is where the decision comes from. No agent can rebuild “why we chose this” from the code. The code shows what was built. The reason sits in someone’s head until it is written down. The article is describing an old problem at a new speed.

The article is also a hypothetical, with no numbers and no named team. I treat it as a sound argument, not as evidence of how often this happens.

Context is the product

I hold that the quality of what goes into an agent decides the quality of what comes out. The code is the easy part. The context is what you have to supply, and nobody else can supply it for you.

Take a clinic that has an agent building its online booking. The owner decided that patients can book without creating an account, because older patients drop out at a sign-up form. That is a business decision with a reason behind it. Suppose the first agent builds the booking form. A second agent is later asked to add reminders and reasons that it needs a verified contact, so it adds a login. A third writes a test that checks login works. Every step passes. The clinic has quietly lost the thing the owner cared about.

The fix is not a smarter agent. The fix is that the second agent was briefed with the decision and the reason. It would then either build reminders around the no-account rule or come back and ask.

Brief, check, then fix the brief

My working loop is brief, produce, check, then fix the brief. Each agent gets the decisions and the reasons, in plain sentences. When the output comes back, I check it against those decisions. A test that confirms login works says nothing about whether booking works without one. The check has to test the decision itself.

When the output is wrong, I look at the brief first. Was the decision missing, vague, or ignored? The article makes a similar point, that each cause needs a different fix. If the brief was missing the reason, I add it. The next agent then starts from a better place.

This is strategy over tactics. If the posture is right and written down, the individual tasks can be delegated. If it is not, no amount of review of individual tasks will catch the drift.

Where this shows up in what I build

In AeroChat, the merchant sets the limits, such as how much it can refund, and the platform enforces them. AeroChat hands anything bigger to the merchant’s team with the order details. The decision lives in the system, so it does not depend on anyone remembering to repeat it.

What a business should do

Pick one piece of work you are handing to an agent this week. Write down the two or three decisions that must still be true when it finishes, and the reason for each. Put them at the top of the brief.

Then write the check before you read the output. If the decision is that customers can book without an account, the check is booking without an account. When the result comes back wrong, change the brief before you change anything else.

Source: e27: AI agents can hand over working code and still lose the decision that made it right