Don’t Let Agents Write Evals from Scratch—They’ll Just Check Whether the Output File Exists
Giving an agent a few good examples and bad examples is more effective than writing long requirements. But bad examples go stale, and examples can also lead to overfitting. In Viv’s discussion, the real challenge is preventing the agent from merely learning to copy the scaffolding.
When you have an agent design evaluations, the most common place things go wrong is the reference point. The model isn’t dumb; the problem is that it doesn’t know what good design means. Viv offered a low-effort approach on X: feed the agent good examples and bad examples that cover design patterns for different tasks, and the Evals and Benchmarks it writes will be much more reliable.
Original post: https://x.com/Vtrivedy10/status/2122089550529151126
Some background is necessary. The design space for evals is huge: how to write the verifier, whether to use a rubric or programmatic checks, how realistic the container environment should be, how to generate synthetic data, and whether task descriptions should be vague or provide enough information. Each of these can steer an evaluation in the wrong direction. Agents amplify context: whatever patterns are in the codebase, they copy. So seed examples give the agent a set of samples to compare against. Viv said they embed the examples they like into skills so the agent has concrete references.
Bad examples may be more valuable. In a reply, Ignacio Ferro said that good examples show only one correct path, while bad examples can name a trap—for example, a judge that rewards length, or a task that leaks the answer. He asked Viv whether they label why each bad example is bad. That question is crucial. Jurly replied that only by labeling the reason for failure can the agent learn design principles; otherwise it just copies surface patterns.
Viv’s answer involves a lot of manual work. It isn’t enough to toss in a few examples. You have to look at which principles the public tasks use, and then check whether those principles can produce tasks that are interesting, not broken, and test something new. He gave two examples: Terminal Bench’s task instructions are very clear—goal, important files, and time limits are all spelled out; when an agent writes instructions from scratch, they are often too long, vague, and leak strategy instead of just stating the goal. Frontier SWE’s optimizer design task shows how to design budgets, counter verifier noise, create automated research tasks, and prevent cheating by modifying files.
Terminal Bench example: https://hub.harborframework.com/tasks/terminal-bench/photonic-waveguide-routing/latest/files/instruction.md
Frontier SWE example: https://www.frontierswe.com/tasks/optimizer-design-v2
There are also finer design choices: when the verifier should use a decision model versus programmatic checks; string matching is fragile, but agents still love to use it. To test visual reasoning, force the harness to read images and videos into context. These details determine whether an evaluation actually measures something.
Counterintuitively, an unseeded agent will easily write a verifier that only checks whether the output file exists. One commenter said plainly in a reply: “unseeded, agents write verifiers that just check the output file exists.” That picture is very typical. Another commenter said much of benchmark progress is curation, not modeling. Bad examples are asymmetric leverage. That judgment is a bit harsh, but it lines up with the phenomenon above.
The risks are there too. Viv warned against overfitting to existing evals; you need to balance stability and diversity. Bad examples also go stale. Mykyta asked whether, once new models stop making that mistake, such examples become useless. Viv said this is inherently a moving target: models get smarter every generation, and evaluations get harder to build. Lachy asked how to find the threshold between stability and diversity. Renan offered an actionable method: copy the harness pattern, then perturb the task assumptions—for example, remove fields that only an oracle would have, or change the tool’s error behavior. If the generated verifier still accepts a plausible but wrong trace, that means it transferred only the scaffolding, not the actual contract.
So seeding isn’t about getting the agent to copy Terminal Bench. Examples need to be labeled with why they’re bad, retired regularly, and used to perturb-test the verifier’s boundaries. The goal is to pass on design principles and contracts. Viv’s dream is to have an idea, message the agent, and have it build the task, test it, and fix it, all automatically. Some domains can already do this; others are still far off. Right now, the more worthwhile investment may be curating a set of qualified bad examples rather than writing another ten pages of prompt.
发布时间: 2026-11-16 09:28