Put the more agentic model in the reviewer's seat

I moved a fleet of scheduled agents from Claude Opus 4.8 to Opus 5, and they got worse at their jobs. Not dramatically. Subtly, in the way that costs you a week before you name it. Answers ballooned. Agents kept changing things nobody asked them to change. Simple tasks grew extra steps. My first instinct was that I had broken a prompt somewhere.

I hadn't. The model was exhibiting exactly the behavior differences Anthropic documents for it. The problem was that I had put it in the wrong seat.

The traits are documented, and they are not bugs

When I stopped treating it as a regression and read Anthropic's own Opus 5 guidance, every symptom was there in plain text. Default responses run longer than on prior Opus models. Task scope expansion: the model can add steps you didn't request, or apply its own judgment about what the task should be. It verifies its own work unprompted, and delegates to sub-agents more readily than 4.8 did. These are documented characteristics of the release, not evidence my prompt was broken. They are the shape of a more agentic model, one that leans toward doing more, checking more, and saying more.

Now put that shape in a bounded, unattended worker that wakes on a timer, runs one narrow task, and goes back to sleep. Those traits can turn into liabilities. In a file-writing worker, scope expansion can mean edits outside the intended task. Unprompted self-correction is a strength by default — but a carried-over "verify your work" instruction compounds it into the over-verification Anthropic warns about, and the tokens add up. More narration means longer runs and more surface for drift. A bounded worker's whole value is that it does the assigned thing tightly and stops, and in my scheduled jobs Opus 5 needed more explicit scope control than 4.8 to hold that line.

The same shape, one seat over, is exactly what you want

Here is the move that took me too long to see. Those traits are only liabilities relative to the role. Point the same model at review instead of building, and each one turns into an asset.

A reviewer that expands scope traces the adjacent code you didn't ask it to look at, and finds the sibling function your change half-broke. A reviewer that verifies unprompted re-derives the claim in your commit message instead of trusting it. Extra detail earns its keep when the brief rewards thoroughness. What breaks a builder builds a critic.

So the fleet settled into an asymmetry that reads backwards on a spec sheet: bounded workers run on Opus 4.8; Opus 5 runs as a reviewer. The worker does what it's told and stops. The reviewer roams, doubts, and over-checks, and I want all three when it's grading someone else's work, and none of them when it's doing mine. None of this says Opus 5 is a weaker model — Anthropic pitches it as strong at exactly this kind of agentic, multi-file coding. It says its disposition fits a different seat than the one I first sat it in.

The remedy that inverts the usual advice

One detail is worth stealing even if you never touch a fleet. Anthropic's fix for Opus 5 over-verifying is to delete the self-verification instructions you carried over from an older model, and they flag that this inverts the usual "ask the model to double-check its work" best practice. On Opus 5 specifically, a carried-over "please verify" can compound the model's own checking and add cost without improving the result. Don't generalize that to every capable model, though: Anthropic's guidance for Claude Fable 5 recommends the opposite, explicit self-verification in long-running prompts. Agenticity alone doesn't set the right verification policy; the specific model does. The same nuance holds for length: verbosity is a prompting lever, not an effort lever — lowering effort reduces thinking volume but doesn't reliably shorten what the model writes.

The honest part

My first write-up of this was cleaner and wronger. It said, in effect, "the new model is worse." I ran it through reviews from two different model families, and they independently challenged that claim: one identified a documented behavior I had misread; the other caught a change I'd made the same week and wrongly blamed on the model. The simple story didn't survive contact with a skeptical reader who wasn't me.

Which is the whole point, twice over. The lesson isn't "newer is worse" — it's that a model has a disposition, and a disposition has a right seat. And the way I found the right version of even this argument was to hand it to the more agentic model and let it do what it does best: expand the scope, distrust the claim, and tell me what I got wrong.