The Computer-Use Agent Kit
AI agents just got hands and cheap hours in the same week. This kit shows you what to hand them, how to brief them, and the guardrails that keep it boring.
What's in the kit
- The field guide — delegate, measure, widen (PDF)
- The delegation shortlist + ROI ladder (XLSX)
- 12 delegation prompt briefs (PDF)
- Guardrails & never-delegate rules (PDF)
Why this kit, why now
Two things blocking real computer-use agents both moved this week. OpenAI shipped GPT-6 Astra with computer use as its hallmark — the model navigates a screen and completes multi-step work like a person would — rolling out to business tiers now. Two days earlier, Anthropic cut cache reads 75%, making always-on agents affordable. Capable hands, cheap hours: the automation window just opened.
What you’ll do with it
- Pick the first task — the shortlist ranks ten tasks against the three tests: can you write the procedure, can you check the result in 60 seconds, is the worst case undo?
- Brief it properly — twelve fill-in briefs built for screen-based work, including the one line that prevents most agent disasters: “if anything is ambiguous, stop and ask — don’t guess.”
- Run the guardrails — spending cap, human gate, sandbox, session limit. Plus the never-delegate list: what stays human, permanently.
- Climb the ROI ladder — one task watched, then unwatched, then widened. The worksheet computes the payback; most first tasks clear the kit price in week one.
The honest fine print
Computer use shipped to business tiers this month — coverage includes first releases, and the tools will improve fast. Agents still misclick, misread, and stall; every unwatched run earns its trust. The guardrails aren’t optional extras — an agent without them is an intern with your passwords and no supervisor.
Who it's for
Operators past the chat-and-drafting phase who want agents doing end-to-end work — without a six-figure automation project.