Post · 28 October 2026
You hired a genius to click Next
Let your main agent write the instructions and send Haiku or Luna to do the clicking.
I think we’re wasting a lot of model effort on browser use and computer use. We give our heaviest frontier model a task, then leave it working through screenshots and tool calls to click buttons and fill forms. That gets slow and expensive for work we can describe as click this, type that, copy this and paste it over there.
We’ve also got very comfortable calling models like Haiku and Luna “lower class”. Two years ago we were impressed by what models could do, and now we talk about these as though following a set of browser instructions is beyond them. I don’t think clicking Next needs our best reasoning model.
In Spending your effort, Thariq Shihipar tried implementing a detailed app specification at different effort levels. The designs and implementations came out much more similar than with a vague prompt, while low effort took 16 minutes and max took 79. That experiment was about building software. My recommendation for routine browser work follows the same idea: work out the instructions first, then spend less effort carrying them out.
Have your main agent write a detailed brief and launch a subagent using the model you choose, with low effort where supported. Give it the steps, what success looks like and when to stop and report a problem.
You can put this in your AGENTS.md, or tell your main agent directly. Replace [MODEL] with the model you want to use, for example Haiku or Luna:
For routine browser-use and computer-use tasks, write detailed instructions and launch a subagent using [MODEL] with the required tools, using low effort where supported. Include success checks and stop conditions, and have it report the result, evidence, blockers and final browser or app state so you can verify the work and send follow-up instructions.
The report matters because the main agent needs enough context to continue. “Done” tells it very little. What changed, what the subagent checked and where it stopped let the main agent verify the result and dispatch the next task without repeating the browser session.
Try it on the next task where you already know the steps. Give the subagent those instructions and have it come back with the result.