State of AI for development: February 2026

Three things from February that we think change decisions, and what we are telling clients to do about each.

Coding agents are extending past the IDE

Cursor shipped full computer use with video output, so you can review what an agent actually did rather than read the diff it left behind. Claude Code added Remote Control for reviewing sessions from a phone, and its unattended session length has grown from 25 minutes to over 45. Devin 2.2 now responds to review comments and fixes CI failures without being asked. The boundary between the inner loop of the IDE and the outer loop of review and deployment is collapsing.

What we are telling clients: the bottleneck is moving from writing code to reviewing what agents did with it. Before extending agent autonomy, invest in the review side: small changes, disciplined CI and enough senior attention to look at what comes back. An agent that can show you its work only helps if someone is resourced to watch.

Agent reliability is lagging capability

METR published work towards a science of AI agent reliability, and the finding that matters is that bigger models are not uniformly more reliable. Scaling improves calibration but can hurt consistency, and autonomous deployment needs somewhere between three and five nines of reliability, which current models are not on track to deliver.

What we are telling clients: draw a hard line between augmentation, where a human reviews the output, and automation, where the agent acts alone, and set a different reliability bar for each. Build an incident-reporting habit around agent failures so you learn from them the way you would from a production outage.

Enterprise agents now arrive with governance built in

Microsoft’s Copilot Tasks executes multi-step workflows autonomously. ServiceNow’s Autonomous Workforce virtualises whole employee roles that inherit enterprise permissions, and its L1 service desk now resolves 90 per cent of its own IT requests. Across February’s enterprise launches, agents arrive with scoped permissions and audit trails on day one.

What we are telling clients: adopt the pattern even if you never buy the products. Any agent you deploy should inherit a defined role with explicit permissions and leave an audit trail, because governance retrofitted after an incident costs far more than governance designed in.

---

This digest is the public edition. Advisory clients get the full monthly review, a tailored edition for their context and the conversation that goes with it.