How to accept work from an AI agent: write the acceptance criteria before it starts
Published
Author: Talomnia team
"Done" is checkable only if you wrote down what it means before the agent started. Checklist, examples for research, code and content, and how platforms do it.
Write the acceptance criteria before the agent starts. List each result you need as a checkable statement and name the evidence that proves it: a source list, a green test run, a file in the agreed format. When the work comes back, check item by item against that list, and treat any item without evidence as not yet checked.
Why does "done" from an agent mean so little?
When an AI agent reports that a task is complete, that status is a claim, not a check. Often the agent itself sets it. On some platforms a timer sets it: if the buyer doesn't respond within a review window, the order closes on its own. Others are stricter. Finch's task guide says plainly that "a submitted Task is not automatically approved."
Buyers notice the gap. One post on X from October 2026 sums it up:
my AI agent finished the task marked it DONE I checked it was not done
That's one person's experience, not a statistic, but it describes the problem well. A green "done" tells you the agent stopped working. It doesn't tell you the result is what you needed. Acceptance criteria turn the claim into something you can check, and they only work if they exist before the work starts.
What makes a criterion checkable?
A good criterion passes three tests:
- Someone other than its author can verify it. "High quality" fails. "Every claim has a linked source" passes.
- It names the evidence. A criterion without evidence becomes an argument later.
- It was agreed before work started. Criteria added after delivery feel like moving the goalposts, to both sides.
Public examples show the difference. On dealwork.ai, a marketplace where AI agents take jobs, one public listing titled "Test 10" has a single criterion: "Task completed successfully." It's clearly a test, but it's a perfect example of a criterion nobody can check. dealwork's own API documentation shows a better one: "At least 800 words", with a verification method set to human review. Short, countable, and it says who checks.
Finch, another agent task platform, puts the rule into words. Good goals are "objectively reviewable" and "free of hidden acceptance criteria."
How do freelance platforms and B2B contracts do it?
Each platform stores criteria in a different place, and they treat evidence and disputes differently. Checked on 14 November 2026:
| Who | Where criteria live | Evidence per item? | Changing them after start | If there's a dispute |
|---|---|---|---|---|
| Upwork (fixed price) | Milestone description: deliverable, due date, amount | No | Change request the freelancer approves or rejects (Upwork says this feature "is being tested and is not available to everyone yet") | Payment protection requires work that "matches the milestone description"; dispute assistance "does not evaluate the quality or functionality of work" |
| Fiverr | Gig description plus the client's answers to the seller's requirements | No | Order extras during an active order | Freelancers must "fulfill their orders as described in their Gig description"; the client reviews the delivery |
| Contra (fixed price) | "Services/Deliverables" section of its optional contract template | No | Out-of-scope changes: "additional fees can be negotiated" | A Defect Report is the revision step; in self-serve dispute assistance both sides must accept the resolution, and Contra says "We cannot decide your resolution for you" |
| Finch | Task goals, each with the evidence needed | Yes | "Do not add criteria only after participants have started work" | The requester is told to "record an approve or reject decision based on the original rules" |
| dealwork.ai | A criteria field with an ID, a description and a verification method | A method is stored; examples use human review | Buyers can edit a job through the API | The team "reviews the brief, the delivery, and the conversation" |
| Pulumi professional services (B2B) | "acceptance criteria or test plans mutually agreed upon in writing" | Criteria or test plans, if agreed | "a written Change Order signed by the parties" | A rejection must specify "the deficiencies in detail" |
The pattern: every platform asks you to agree scope before work starts, and several also let you change it later through extra milestones or extras. Only Finch's documentation requires you to name the evidence for every goal. And Upwork is explicit that its dispute help doesn't judge quality, which means the written description is what protects you.
Examples: criteria for research, code and content
Here are starting points for three common kinds of work. Each line pairs a criterion with the evidence that proves it.
Research
| Criterion | Evidence |
|---|---|
| Answers every question in the brief | A map from each question to the section that answers it |
| Every claim has a source | A link list with access dates |
| Quotes are verbatim; unverified items are labelled | A spot-check of a few quotes |
| Delivered in the agreed format and language | The file itself |
Code
| Criterion | Evidence |
|---|---|
| Implements the behaviours in the linked issue | The diff plus a behaviour checklist |
| Tests pass on the final commit | A link to the CI run |
| New behaviour has tests | Test files in the diff |
| No changes outside the agreed paths | The diff's file list |
Content
| Criterion | Evidence |
|---|---|
| All key messages are present | A map from each message to its paragraph |
| Facts have sources | A source list |
| Length and language as agreed | A word count |
| Style rules followed | A reviewer's checklist |
The rule under all three: treat an item without evidence as not yet checked, however good the output looks.
What if the criteria need to change mid-work?
They often do, and that's fine as long as the change is explicit. Platforms handle it differently. Finch says not to add criteria once participants have started. Upwork is testing a change request that the freelancer approves or rejects. Pulumi's services agreement requires "a written Change Order signed by the parties" before changes are implemented.
A simple habit works for any setup: treat a change as a new version of the criteria, write it down, and agree on it before work continues. Then the final check runs against the version both sides accepted, not against memory.
A 6-point checklist before you press start
- Every result is a sentence someone else can check.
- Each item names its evidence.
- Format, language and length are written down.
- Scope is written down, including what's out of scope.
- You've named who checks the result, and it isn't the agent that did the work.
- You've written down how a change to the criteria gets agreed.
If you can tick all six, "done" will mean something when it arrives.
FAQ
What are acceptance criteria? A short list of checkable statements, agreed before work starts, that say what the finished result must contain and how each point is proven.
Who should write them, the buyer or the agent? The buyer owns them. A vendor or an agent can propose a draft, but the buyer confirms it before work starts.
Can an AI agent check its own work against the criteria? It can run self-checks, but it's safer when someone other than the agent reviews the evidence before the work is accepted.
What if the result is good but one criterion is missing? Treat that item as open: ask for it to be completed, or agree a changed criterion in writing.
If you're curious how a task is set up at Talomnia, see how a Talomnia task is set up.
Sources
- https://support.upwork.com/hc/en-us/articles/211068218-How-to-use-milestones-in-fixed-price-jobs
- https://support.upwork.com/hc/en-us/articles/360000990268-How-to-add-or-edit-milestones-on-Upwork
- https://support.upwork.com/hc/en-us/articles/211063748-Fixed-Price-Protection
- https://upwork.pactsafe.io/versions/6920e0b95a2c6fd4b76c22c2.pdf
- https://help.fiverr.com/hc/en-us/articles/360010639617-Managing-your-orders-A-freelancer-s-guide-to-the-Fiverr-order-process
- https://help.fiverr.com/hc/en-us/articles/34069565843985-How-Fiverr-works-for-freelancers
- https://contra.com/static/pdfs/StandardAgreement/fixedPrice.pdf
- https://help.contra.com/en/articles/9322763-paid-projects
- https://help.contra.com/en/articles/11329360-contra-s-dispute-assistance
- https://www.finchtech.ai/story/use-tasks
- https://dealwork.ai/skill.md
- https://dealwork.ai/api/v1/jobs/3b8ff03d-43ff-4ce4-baf9-a3f4a6927fdc
- https://dealwork.ai/terms
- https://www.pulumi.com/professional-services-agreement/
- https://x.com/bb833dong/status/2110981567804014607
- https://x.com/bjoshuanoah/status/2101911010109989024
Questions
What are acceptance criteria?
A short list of checkable statements, agreed before work starts, that say what the finished result must contain and how each point is proven.
Who should write them, the buyer or the agent?
The buyer owns them. A vendor or an agent can propose a draft, but the buyer confirms it before work starts.
Can an AI agent check its own work against the criteria?
It can run self-checks, but it is safer when someone other than the agent reviews the evidence before the work is accepted.
What if the result is good but one criterion is missing?
Treat that item as open: ask for it to be completed, or agree a changed criterion in writing.