Algorithm & Blues

Anthropic’s latest research looks at telemetry from deployed agents, and the findings offer a grounded view of how autonomy is developing in day-to-day workflows.

The researchers used telemetry from Claude Code sessions and API tool activity to track how long agents run without interruption.

At the tail, session duration at the 99.9th percentile nearly doubled between Oct 2025 and Jan 2026, going from under 25 minutes to more than 45 minutes.

Benchmarks were fairly steady, but behavior shifted; people let agents run longer and finish more work before stepping in.

Autonomy in this setting reflects a mix of capability, product design, safeguards, and user comfort with delegation. As those factors mature, workflows that previously required close supervision begin to run with less interruption.

Most usage remains routine; activity at the tail is where early value and failure patterns appear first.

The findings highlight a visibility gap around delegated autonomy and how clearly organizations can observe it.

Several observable indicators are worth noting:

• unattended execution duration • reversibility of agent actions • approval checkpoints and escalation paths • tool invocation patterns and blast radius • exception handling and recovery behavior

In practice, autonomy functions as a runtime property that can be observed and constrained alongside other operational signals.

Anthropic describes a related dynamic as deployment overhang; capabilities are advancing faster than delegated autonomy, creating tension across architecture, controls, and operating models.

Agent progress will be defined not only by model capability, but by how deliberately autonomy is granted.

Paper: https://lnkd.in/eiFkEQNH

hashtag #AI hashtag #AgenticAI hashtag #AIGovernance hashtag #EnterpriseAI hashtag #AlgorithmAndBlues