Computer Use Gets Faster. Control Matters More.
Astra can work across browsers and desktop software. The workflow still needs scoped access, visible actions, exception handling, and human approval before consequences leave the workspace.
Computer Use Must Separate Preparation From Authority
Astra can navigate interfaces and complete multistep tasks faster. The control problem is deciding which actions it may prepare, which it may execute, and which must stop for approval.
OpenAI reports a new frontier in computer use for Astra, including forms, CRM updates, calendar work, research, documents, software installation, scientific tools, and frontend QA.
GPT-6 Astra is not yet available in Buda. The goal here is to map the control plane a computer-using agent needs, not to imply that Buda can run Astra today.
What Do the Computer-Use Results Show?
Astra improves both task success and elapsed time in OpenAI's reported evaluations. Production performance will still depend on the interface, harness, permissions, network, and confirmation policy.
| Evaluation | GPT-6 Astra | What it measures or comparison |
|---|---|---|
| OSWorld 2.0 | 72.6% | 65.7% for GPT-5.6 Sol |
| OSWorld task time | About 40 minutes | About 75 minutes for GPT-5.6 Sol; roughly 47% less |
| ScreenSpot-Pro | 92.7% | Visual target localization without tools |
| Agents' Last Exam | 59.3% | Professional tasks in real software |
| Mind2Web harness | 1.9x faster completion | Astra plus OpenAI's updated Codex harness |
Source: OpenAI, GPT-6 Astra: A new generation of intelligence. Verified September 4, 2026. Scores are specific to OpenAI's reported evaluation setups.
A Seven-Step Controlled Computer-Use Workflow
The agent should never infer broad authority from access to a browser or desktop. Scope and consequence determine the next action.
Define the allowed surface
Name the domains, applications, accounts, records, and action types the workflow may touch.
Artifact: Scope policyValidate the inputs
Check required fields, source freshness, identity, and missing information before opening the interface.
Artifact: Input packetObserve before acting
Read the current screen and compare it with the expected state; stop on an unexpected layout or permission prompt.
Artifact: State capturePrepare the action
Fill fields, assemble the update, or stage the operation without submitting an irreversible change.
Artifact: Action previewVerify the preview
Check totals, recipients, dates, permissions, and any side effects against the source data.
Artifact: Preflight resultRequest approval when required
Present the exact consequence and evidence to the accountable operator.
Artifact: Approval artifactAct and record the result
Execute only within granted scope, then capture the final state and any exception.
Artifact: Action logThree Failure Modes That Benchmarks Do Not Remove
Higher model capability changes the risk profile; it does not remove operational failure.
Interface drift
A button moves, a dialog changes, or the active account differs from the expected one, but the agent continues by analogy.
Accidental commitment
Drafting and submission happen in one step, leaving no clean point for a person to inspect the consequence.
Hidden retries
The agent repeats a failed action and creates duplicate records, messages, bookings, or charges.
What the Buda Control Plane Should Hold
Keep context, execution evidence, and the review decision in one persistent workspace.
Scoped access
The allowed apps, domains, records, credentials, and action types for this workflow.
Visible execution
Screenshots, current state, staged values, retry count, and the exact point where the workflow paused.
Confirmation policy
Rules for read, prepare, submit, send, purchase, delete, and any action that crosses an authority boundary.
Computer-Use Review Checklist
Questions About GPT-6 Astra Computer Use
What can GPT-6 Astra do on a computer?
OpenAI shows Astra working with browsers, forms, CRM systems, calendars, documents, scientific software, installation tasks, and frontend QA.
Is Astra computer use available in Buda?
Not yet. Buda has not announced GPT-6 Astra availability.
What does 72.6% on OSWorld mean?
It is OpenAI's reported partial score on an offline OSWorld 2.0 subset under the evaluation setup described in the release.
Can a computer-use agent submit forms automatically?
It can technically act, but consequential submission should be controlled by permissions and explicit approval policy.
What should happen when the interface changes?
The agent should stop or escalate instead of guessing, preserve a screenshot, and describe the mismatch for review.
What should a computer-use audit trail include?
The requested goal, granted scope, source inputs, screenshots or action evidence, exceptions, approvals, and final result.
Let the agent operate the interface. Keep authority visible.
Start from the workflow, evidence, and review boundary. Model access comes after Buda officially supports Astra.