From my perspective balancing AI Agents Agency with Control is one of the most important themes for

From my perspective balancing AI Agents Agency with Control is one of the most important themes for 2026. We need to get this right both as builders and users for AI Agentic systems.

Anthropic’s study “Measuring AI agent autonomy in practice”[1] nicely fits into this as they started to study autonomy of AI agents from two data sources: full Claude Code sessions (how long Claude runs before stopping, auto-approve and interrupt rates) and individual tool invocations on their public API. The study notes it uses “autonomy” somewhat informally and analyzes API actions in isolation, so these are proxies, not a settled measure. Already this first iteration provides some nice insights. Obviously a high focus on Coding use cases (software engineering is nearly 50% of API tool calls), but also indicating a wide variety of other use cases which resonate with my experience from customers I’m working with.

Further insights how Anthropic observed their measured autonomy — for Claude Code, the per-turn working time before stopping — varying over time and across model versions.

Again a reminder of everyone building and deploying AI Agents to not forget to 1/ build in observability to be able to understand what their agents are doing and how they are used and then 2/ actually use this data to derive insights.

What is your take on this?

Sources [1] McCain, Millar, Huang, et al. (Anthropic), “Measuring AI agent autonomy in practice”, 2026-02-18 — https://www.anthropic.com/research/measuring-agent-autonomy

• AI
• AIAgents