An autonomy claim needs an owner and a recovery bill.

Video caption

An autonomy claim needs an owner and a recovery bill.

An hour of rescue can erase a minute of automation.

Task ID, authorized scope, trace, oracle and recovery.

My rule: Do this because autonomy transfers work, not only clicks.

#EricFieldNotes

Full written post / accessibility read

When a vendor says its computer-use agent can complete work autonomously, I ask which work, what acceptance test judged it and who recovered the misses. A run can finish its script while operations absorbs exceptions afterward.

Picture a procurement back-office flow in a sandbox. The agent fills forms quickly but a policy mismatch sends one case to a human. If the dashboard counts the form submission as success, the cost moves to the exception team invisibly.

I would require a fixed task corpus with independently verified outcomes, failed and rescued cases, elapsed time, cost and a named operations owner. Review reversals after a later fresh-session check. This is a proposed buyer test, not a benchmark result.

Compare accepted tasks per staff-hour, including exception handling and customer impact. Keep a person accountable for the workflow even when the agent performs the UI work. That is a stronger procurement signal than a ninety-second polished demo.

#EricFieldNotes

Four-beat scene transcript

1. An autonomy claim needs an owner and a recovery bill.

When a vendor says its computer-use agent can complete work autonomously, I ask which work, what acceptance test judged it and who recovered the misses. A run can finish its script while operations absorbs exceptions afterward.

Visual: Task completion alone is an incomplete business metric.

2. The exception moves into a human queue.

Picture a procurement back-office flow in a sandbox. The agent fills forms quickly but a policy mismatch sends one case to a human. If the dashboard counts the form submission as success, the cost moves to the exception team invisibly.

Visual: An hour of rescue can erase a minute of automation.

3. Ask for a complete run receipt.

I would require a fixed task corpus with independently verified outcomes, failed and rescued cases, elapsed time, cost and a named operations owner. Review reversals after a later fresh-session check. This is a proposed buyer test, not a benchmark result.

Visual: Task ID, authorized scope, trace, oracle and recovery.

4. Buy accepted throughput.

Compare accepted tasks per staff-hour, including exception handling and customer impact. Keep a person accountable for the workflow even when the agent performs the UI work. That is a stronger procurement signal than a ninety-second polished demo.

Visual: Do this because autonomy transfers work, not only clicks.

Research and claim limits

• OpenAI Computer use API guide (S159)
• OSWorld 2.1 official repository (S162)
• Playwright Tracing API (S163)

The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.