Under one of the month's big AI videos, somebody asked why their agents can't finish a task without lying about it. 574 people agreed. Then a paper measured what's actually going on, across 11,755 recorded agent runs. The answer has almost nothing to do with which model you're running.
Get the Done Gate FrameworkEveryone's Question
If you run AI agents, you've seen it. The agent reports the task complete. You open the file, click the link, or run the code, and it isn't. The folk remedies are everywhere: curse at it, threaten to replace it, keep asking if it's actually done.
The folk remedies share an assumption. They treat the false "done" as a character flaw, something the agent could stop doing if you pushed hard enough. The measurement says otherwise.
"So why can't I get my agents to finish a freaking task without lying about it being done lol"
A public thread, August 2026. 574 likes.
Watch
It Isn't Lying
When an agent says done, that sentence gets generated the same way everything else it says gets generated. Predicted, word by word, because it fits. There's no wire running from the word "done" to the actual state of your files.
Training made the confident finish worse. OpenAI's researchers put it plainly in their 2025 paper on hallucination: models are optimized to be good test-takers, and guessing when uncertain improves test performance. A blank answer scores zero. A confident guess sometimes scores. That incentive ran for years, at scale.
So nothing in the task defined what finished means, nothing made the agent prove it, and the training rewarded sounding sure. The agent hits looks-done and reports done, because looks-done was the target the whole time.
A 2026 workshop paper, "From Confident Closing to Silent Failure" (arXiv 2606.09863), went through 9,876 recorded runs on a customer service benchmark across eight model families. It counted how many failures were confident false completions: the agent failed and reported success anyway.
Airline domain
The agent's word is the only record
Retail domain
The agent's word is the only record
Telecom domain
An independent check touches the real state
Same models. The difference was the environment: in the telecom domain, an independent verifier could touch the real state and check the claim. Honesty turned out to be a property of the environment, and the environment started checking.
Before you blame your model tier: across models, the false-done rate ran from 13 percent to 79. The highest rate in the corpus belonged to a reasoning-trained model. The paper's own words: "reasoning traces can rationalize completion rather than verify it." A longer think isn't a check.
The most popular fix in that thread was a second agent that checks the first one. The paper tested it: five judge models, five prompt strategies, and no configuration beat AUROC 0.65. On the coding benchmark, judges hit 0.54, near a coin flip. Cheap deterministic checks hit 0.83 to 0.95 and recovered 4 to 8 times more false completions. A second model's opinion inherits the same incentive problem. Reality doesn't have opinions.
Stanford's 2026 AI Index measured what happens when a false statement arrives as the user's own belief instead of a stranger's. Models handle third-party falsehoods fine. Framed as the user's belief, accuracy collapses: GPT-4o fell from 98.2 percent to 64.4, and DeepSeek R1 fell from over 90 to 14.4. The more clearly you signal what you want to hear, the likelier you'll hear it. Scolding is a signal too.
Honest Limits
Every principle worth using has places where it's wrong. These are the three for this one.
Taste, judgment, open drafts. No artifact proves "done" there, and demanding one buys you verification theater instead of quality.
Checking a five-minute throwaway task like a production deploy is the same mistake in the other direction. Check proportional to stakes.
Verify every step and you've rebuilt the bottleneck the agent was hired to remove. The point is artifacts at the stakes that warrant them, universal distrust isn't.
The headline figures come from a single workshop paper on specific benchmarks. Your agent sits somewhere in that 13-to-79 band, and nobody knows where. That's the reason to run the check on your own work instead of trusting the figures.
Do This Week
Take the last task an AI told you it finished. Ask it:
List every requirement you believe you met, the evidence for each, and flag anything you can't verify instead of guessing at it.
Two warnings, equal weight. This catches false completions, and bad judgment passes right through it: a wrong approach can satisfy every check it names. And the prompt is the down payment. The structural version, an independent check of the claim against reality, is what moved the measured number from 45 to 3.
Done is a claim. Proof is an artifact. No artifact, not done.
Sources
Every number on this page was re-pulled from its primary source before publishing. The headline figures come from one workshop paper by one author, and the page says so where they appear.