The hardest line in any agent is the one that says “done”. Too eager and it stops with the job half-built; too stubborn and it burns your budget on the same mistake. Here is how Oppermind agents decide, and why the checks get written before the work.
In an AI agent is a loop we said the line most people leave vague is the last one — check(), the thing that decides whether to go round again. “Until the task is complete” is not a stop condition. This article is the answer to that: what a real stop condition looks like, and how it is built into Agents mode.
Most of what you ask an agent is ordinary: summarise this, draft that, pull the figures out of the other thing. It runs, it answers, it is over. No goal is opened, because there is nothing to carry through.
The other kind is the request that has to end in a state: build the storefront and give me a working preview; produce the three-year model with a chart; find whether this claim holds. You either say so — “carry this through to completion” — or mark the problem as hard, and the agent opens a goal. Everything below is what happens inside one.
Before the agent touches the work, it writes the acceptance checks. For a storefront: the preview URL resolves; the catalogue shows twelve products; the pre-order form submits and captures an email; the health checks pass. For a model: the assumptions sheet exists; every downstream figure references it; the chart shows revenue against cash; the break-even month is present.
Why first? Because a check written afterwards is just a description of whatever got produced. Written first, it is a commitment the work can fail. This is the single most important design decision in the whole mechanism, and it is the one that separates an agent that reports success from one that can be wrong in a way you can see. Visibly wrong is far more useful than confidently finished.
When an attempt completes, it is graded against the checks — and wherever it can, the grader reads the actual deliverable rather than the agent’s account of it. It opens the file. It fetches the preview. It counts the products. The agent does not get to mark its own homework by saying it did the homework.
Each check comes back with a verdict and an evidence label, so when you read the grading you can see not just pass or fail but what the grader looked at to decide. That matters at the end, when the handover has to be trusted.
Anyone who has watched an agent retry knows the failure mode: attempt two is attempt one again, delivered with more confidence. Goals keep a ledger of every approach tried, and an approach in the ledger is never repeated. When the obvious approaches have failed, the agent is handed deliberately different ones to try — not a variation, a different route.
The practical effect is that a goal’s attempts look like a person working a problem rather than a person hitting refresh.
A goal has a budget you set. It ends in exactly one of two ways: the checks pass, or the budget is spent. There is no third ending where it declares victory because the last attempt felt close.
When the budget runs out, the goal closes on the best result so far and hands over with a plain statement of what is still missing. You can also close it early yourself — open the session and say finish with what you have and list what is missing — and the handover has the same shape. That shape is the point: you always get the work and an honest account of its gaps, never one without the other.
In the shoe-store walkthrough, the last message to the team was: grade what you built against the checks you wrote before you started, and give me one plain list of what is still missing before this store can take its first paid order. The answer was payments, supplier agreements, an ABN and GST, a returns policy, an ads account and a domain.
That list is only possible because of the four steps above. Checks written first gave the agents something to fail. Grading against the real files made the verdicts mean something. The ledger meant each agent had actually tried alternatives before conceding. And the best-so-far handover meant conceding did not throw the work away.
| Mechanism | What it does | When you meet it |
|---|---|---|
| Needs you | The agent stops to ask a question or wait for approval, typically before something irreversible: sending, publishing, changing a record. Shows as a status in the Agents list and as Approve / Reject in the chat. | Any run that reaches a gate you set, or a question it cannot answer alone |
| Routines | An agent on a clock. Every Monday at 8am, every hour, whenever you set. Runs, leaves a real artefact, stops. | Recurring work that should happen without you typing |
| Listeners | An agent on a trigger. Runs when an event lands rather than when a time arrives. | Work that should happen because something changed |
| Known results | Published values or test vectors found by research are handed to a check as ground truth. A verification checks them directly, and findings are reported as reproductions of known results or as unverified until independently confirmed. | Research and verification goals where a wrong number is expensive |
Routines and listeners are the two clocks of an agent company: one runs on time, one runs on events. Needs you is the management layer. Known results is what stops a confident-sounding number from being treated as a fact.
The mechanism does the grading. You still decide what “done” means, and a goal is only as good as the state it names. Three habits:
Write those three and the agent’s own acceptance checks will be sharp. Leave them out and the checks will be as vague as the request.
Goals, the ledger, the budget, best-so-far handover, Needs you, routines, listeners and known results are all part of Agents mode in the Oppermind workspace today, alongside group chats and the deliverables landing as real files. The Academy’s build your first AI worker is the eighteen-minute hands-on version; the workers reference is the short one, and its first rule — start with work that can be reviewed before it has an external consequence — is the same rule this article has been circling.
Free with no card, then Starter A$9.95, Pro A$29.99 and Pro Plus A$59.99 a month. Unlimited documents, spreadsheets, decks and designs on every plan.
Whatever you’re here to make, make more of it.
A check is only as good as the thing it can measure. “The preview resolves” is a strong check because it is either true or not. “The brand voice is right” is a weak one because a grader reading a file cannot settle it. Goals are excellent at the first kind and honest about the second — they will report the check as met on the evidence available, and the evidence label will tell you how thin that evidence was. Read the labels. The mechanism removes a lot of ways to be fooled; it does not remove the need to look.
A goal is what an agent opens when you ask for something to be carried through to completion, or mark a problem as hard. Before any work starts, the agent writes the acceptance checks. Every attempt is then graded against those checks by reading the real deliverable where it can. The goal ends only when the checks pass or the budget you set is spent, and then it hands over the best result with a plain statement of what is still missing. Ordinary requests never open a goal.
Because a check written afterwards is a description of whatever got produced. Written first, it is a commitment the work can fail. This is the difference between an agent that reports success and one that can be wrong in a way you can see — and being visibly wrong is far more useful than being confidently finished.
A record of every approach the agent has tried on a goal, kept so that no approach is repeated. When the obvious approaches fail, the agent is handed deliberately different ones to try. It is how a second attempt is genuinely a second attempt rather than the first one again with more confidence.
The goal closes on the best result so far, with a plain list of what is still missing. It does not pretend the checks passed. You can also close it early yourself by telling the agent to finish with what it has and list what is missing — the handover is the same shape either way.
The agent has stopped to ask a question or wait for an approval — typically before something that cannot be undone, like sending, publishing or changing a record. It shows as a status in the Agents list and as a message with Approve and Reject in the chat. Open it and answer; the run continues from there.
check() line made concrete.This article describes Oppermind Agents mode as it ships on 13 September 2026; behaviour is subject to change with updates. Acceptance checks, grading and evidence labels reduce but do not eliminate the need to review agent output before relying on it, and you remain responsible for any decision or action taken on the basis of an agent’s work. This article is general information about software, not advice. Oppermind platform prices are current as at 13 September 2026, are in Australian dollars, and are subject to the plan terms at checkout.
Goals with acceptance checks, a ledger, a budget, and a handover that tells you what’s missing — in every plan.