Key facts
| Router model | plugsky-fusion escalates per request across tiers (live) |
| Agents | Memory, tools and orchestration are live |
| State format | Prefer DOM or accessibility text over screenshots when the target exposes it |
| Loop controls | Cap steps, retries and per-step context in your orchestrator |
| Pricing | Flat monthly self-serve plans with no per-token charges on self-serve |
| Free tier | plugsky-micro and plugsky-lite on the free plan, no card required |
| Roadmap | The images endpoint is coming soon; validate multimodal input per model today |
| Governance | Scoped keys, RBAC and audit logs for agent actions |
TL;DR
- Every screenshot-heavy step is expensive; prefer text state when you can.
- Route clicks, typing and navigation to cheap tiers.
- Escalate planning, validation and error recovery to strong models.
- Cap total steps and retries per task, then hand off.
- Cache stable page state instead of resending it every step.
How it works, step by step
- Map the task into steps and mark which ones are mechanical versus judgemental.
- Extract page state as DOM, accessibility tree or form data instead of screenshots wherever the target allows.
- Assign cheap tiers to clicks, typing, scrolling and simple validation.
- Escalate plan revision, anomalous screens and failure recovery to stronger models.
- Cap steps per task and retries per action, with a human handoff when exceeded.
- Cache stable regions of page state within a run and drop them when they change.
- Measure step count, recovery rate and completion per task type before widening autonomy.
Try it yourself
Open the agent workflow designer →
Screenshot loops are the cost centre
A computer-use agent typically observes a screen, decides an action, acts and repeats. If every observation is a full screenshot, each step carries image data and often the entire history besides. Long tasks compound that, and retries replay it. The result is one of the most expensive agent patterns per completed task.
Reduce observation size first. When the target application exposes structured state — DOM, accessibility tree, form values, API responses — send that instead of pixels. It is smaller, more precise for clicking and typing, and cheaper to process. Screenshots remain useful for layout-dependent decisions and visual verification, but they should be the exception, not the default. Note that the images endpoint is coming soon, so validate multimodal input per model in your own environment.
Tiering steps without losing reliability
Most computer-use steps are mechanical: locate a field, type a value, submit a form, verify a confirmation banner. These run well on cheap tiers. Judgement concentrates in planning the task, recovering from unexpected screens, and validating that the outcome is correct — those deserve strong models.
- Cheap tier: element location, form entry, navigation, simple success checks.
- Strong tier: plan revision, ambiguous screens, error recovery, final validation.
- Pinned paths: payment, deletion and external communication actions, regardless of tier.
- Human gate: irreversible actions with real-world consequences.
Budgets, caching and evaluation
Enforce a step budget per task and a retry budget per action in your orchestrator, and treat budget exhaustion as a normal outcome that hands off to a person. Cache stable page state within a run — a header, a navigation bar, a session banner — and invalidate it when it changes. Never append an unchanged screenshot or DOM snapshot to the next request.
Evaluate per task type with completion rate, steps taken, recovery rate and human interventions. A cheap tier that completes the same task in the same number of steps is a real saving; one that adds five retries is not. Start on the free plan with plugsky-micro and plugsky-lite, then use the 14-day full-access trial for stronger models on hard recovery cases; plan details are on the live pricing page.
Honest comparison
| Cost lever | Routed computer-use agent | Frontier model every step | Screenshot-only cheap model |
|---|---|---|---|
| Observation | Structured state first | Images plus full history | Images, often misread |
| Mechanical steps | Cheap tiers | Frontier price per click | Cheap but error-prone |
| Recovery | Escalated with context | Native strength | Usually fails |
| Step budget | Caps and handoff | Often unbounded | Often unbounded |
| Evidence | Completion and steps per task | Token totals only | Token totals only |
Frequently asked questions
Why are computer-use agents the most expensive pattern?
They combine long task histories, frequent observations and retries. Each step may carry screenshots and state, so dozens of steps produce a large aggregate request volume.
Do I need images for computer use?
Not always. DOM, accessibility trees and form data are usually more precise and cheaper for clicking and typing. Reserve visual observations for layout-dependent or verification steps, and note that the images endpoint is coming soon, so validate multimodal input per model.
Which steps can run on cheap models?
Element location, typing, form submission, navigation and simple confirmation checks. Planning, unexpected screens and failure recovery should escalate to stronger tiers.
How do I stop an agent from clicking forever?
Enforce step caps per task and retry caps per action in the orchestrator, and treat exhaustion as a handoff rather than a failure to retry indefinitely.
How do I handle irreversible actions?
Pin them behind explicit validation and, where the consequence is real, a human approval gate. Tier choice does not substitute for a safety gate.
Can I cache page state?
Yes. Cache stable regions within a run and invalidate on change. Never resend unchanged snapshots to the model; that is pure cost with no informational value.
How do I evaluate savings?
Track completion rate, steps per task, recovery rate and human interventions by task type. Efficiency claims without task outcomes are meaningless for agents.
Can I prototype on a free plan?
Yes. plugsky-micro and plugsky-lite are on the free plan with no card, and the 14-day full-access trial is available for stronger models.