TL;DR - Anthropic have updated Claude Code to change token-heavy /verify, /code-review and /deep-research skills manual-only. Agents can no longer invoke these, rationing autonomy. Think of autonomy as a budget that you have to allocate.

So you’ve spent time adding hooks, wiring up skills. Your agent verifies its own work without being asked. Then Anthropic push an update. You read the changelog and get a sinking feeling. They’ve an update turning it all off, on purpose.

What changed: recent Claude Code releases made /verify, /code-review, and /deep-research manual-only. The agent no longer decides to run its most expensive behaviours, so now you have to. Whilst it’s obvious why - these are especially expensive operations and Anthropic are losing money - it still represents a regression in your productivity.

The search cap: in addition to ‘manualising’ certain skills, there’s a session-wide WebSearch limit (default 200) to kill runaway search loops. This is to stop runaway agents googling themselves into oblivion. This one is not remotely egregious: what new information could an agent reasonably gain from the 201st web search?

Read more on the Claude Code changelog and release summaries.

In all honesty, it’s probably a decent move in the aggregate. What percentage of auto-triggered deep research or code review is actually required? We’ve all had agents run off on research tangents or take 10 minutes to review whole codebases following cosmetic changes.

Couple this with the trust angle: would you prefer the agent demands permission to do long, token-intensive tasks? Definitely if you’re on the API, and probably if you’re the kind of developer that sails close to the limits of your subscription. Let’s not lean too heavily into this sentiment, though. It is a budgetary consideration, not a UX one.

What we do lose however are the times when Claude Code genuinely, autonomously uncovers things that will help us. You might not consider deep-research essential for what you’re up to but it could be the only thing to surface unknown unknowns in your work.

Autonomy, therefore, is a budget consideration first and foremost. I still believe we’re in the early-Uber era of AI where vendors are heavily subsidising our usage. Remember when Uber first started and you could get across town for half the incumbent taxi fee? Now Uber’s more expensive. Expect the same for tokens once your organisation is fully-dependent on it.

There are now two questions to ask for every hook/skill/tool:

In one of the companies we work with, Dagenta, I built a skill for backtesting price forecasting models which iterates over hundreds of GB of data in S3. Developing new models with Claude sometimes triggered this skill, adding hours and tokens onto my prompt completion. In such circumstances, you definitely want to be asked.

A crude mapping of task vs automation should therefore be this:

Our free tool this week lets you run an audit on every skill, hook and MCP tool your agent can trigger to see which ones should be auto and which should be manually gated. It outputs an analysis with three columns: name, auto or manual and cost-of-wrong.

Our rule-of-thumb here is that anything taking more than a minute and/or touching anything external should always begin life as manually gated until you feel it’s earned the auto status.

You’ll find this as a new skill called /autonomy-audit in the temrel-agentic-toolkit.

Add it as a skill (run manually, naturally) and it will output an autonomy-audit.md file in your repo’s root. It makes no changes, only reporting what you might consider changing.

The industry narrative right now is a straight line to full autonomy, but the people with the most usage data (who are also losing money at a vast rate) just drew a line under it. Expect vendors to ship more and more constraints. You’ll need to be a little trickier to truly have autonomous coding agents.

Your edge isn’t maximum autonomy therefore, it’s more calibrated than that. Know which behaviours have the right to fire on their own.

Which agent behaviour have you deliberately taken back to manual, and what did it cost you before you did? Hit reply, I read everything.

See you next week.