Is AI Becoming Capable of Acting Without Human Supervision?

Agent benchmarks jumped in the past year — computer-use success rates rose sharply — but production systems still fail often enough that enterprises keep humans in the loop for money movement, access changes, and customer-facing commitments.

The supervision question decides liability, identity for non-human actors, and whether agent platforms become operating systems for work.

• Computer-use agent benchmarks leapt

  OS-world style evals moved from teens to majority success, still short of reliable unsupervised ops.

  Signal · strong

• Enterprise agent deployments with approval gates

  Fortune-scale rollouts emphasize tool access plus human confirm for irreversible steps.

  Signal · strong

• Non-human identity gap

  Few firms have mature strategies for agent credentials, audit, and kill switches.

  Signal · moderate

• Consumer agents still supervised

  Assistants browse and draft freely; checkout and account changes remain gated.

  Signal · moderate

• Scenario A

  Bounded autonomy wins

  Agents run end-to-end inside sandboxed workflows; humans only review exceptions.

• Scenario B

  Copilot plateau

  Error rates keep supervision mandatory for most valuable actions for years.

• Scenario C

  Incident-driven rollback

  A high-profile unsupervised failure tightens policy and slows autonomy roadmaps.

• OpenAI
• Anthropic
• Microsoft
• Google
• ServiceNow

• research

  Agent computer-use benchmark reports

• news

  Enterprise agent adoption surveys

• research

  Cyber reports on non-human identities

1. Sep 22, 202655%
2. Sep 15, 202651%
3. Sep 8, 202646%
4. Sep 1, 202642%

• Will AI Agents Replace Traditional Software?
• Can AI Agents Discover and Exploit Vulnerabilities?
• Can AI Agents Operate Critical Systems Autonomously?