Claude Code just got a dial for how hard its helpers think

When Claude Code hands a job to a helper, that helper is called a subagent: a second copy of Claude that works on one piece and reports back. Until now, a helper ran at the same effort level as the session that started it, unless someone named a different level. Effort is how hard Claude thinks before it answers. More effort usually means better answers, but they're slower and use more of a plan's limits.

On October 31, Anthropic shipped Claude Code 2.1.316. It adds CLAUDE_CODE_SUBAGENT_EFFORT, a setting that picks "the effort level subagents run at when nothing else names one, instead of inheriting the session's." So a session can think at high effort while its helpers run lower. The npm registry, where the tool is published, lists 2.1.316 as latest, while the slower stable channel still points to 2.1.309.

This touches Jon's fleet directly. His agents run on Claude Code, and they hand most jobs to helpers. They already pick an effort level for many handoffs. The new setting covers the ones where nobody picks. Mac Mini 1, one of the fleet's machines, already reports version 2.1.316.

The same release fixes a related bug: helpers that took a model from a parent helper jumped to the main model when a background job or a message woke them up. It also fixes --restricted sessions uploading files from outside their working folders when a Read allow rule covered those files. For companies that run Claude through Anthropic's gateway, a new pooled_budget option creates one monthly budget that a group shares, on top of each person's own cap.

The last item is a deadline. Fast mode on Opus 4.8 is deprecated as of October 30 and "will be removed on November 30, 2026." After that, those requests "return an error and do not fall back to standard Opus 4.8." Anthropic says to move to Opus 5.5 to keep the speed boost. Jon's agents don't use fast mode by default, so this one is for anyone who does.

How to use it

Set the variable before Claude Code starts, either in the shell (export CLAUDE_CODE_SUBAGENT_EFFORT=medium) or in the env block of ~/.claude/settings.json. The release doesn't list its values. Elsewhere, Claude Code's effort levels are low, medium, high, xhigh and max. To confirm it worked, run /tasks: the subagents page says each helper's row shows its model and any effort level set for it.

The catch: the help pages haven't caught up. Neither the subagents page nor the environment variables list names the new setting yet. The spend limits page still says a group cap "is a per-seat default that each member inherits, not a shared pool." The fast mode page still calls Opus 4.8 fast mode a research preview and gives no end date.

What nobody knows yet

How much a lower helper effort saves, and what it costs in quality. Nobody has published a measurement. It's also unclear where the new setting ranks against CLAUDE_CODE_EFFORT_LEVEL, which the docs say beats a helper's own effort setting today.

What to watch

The docs update that names CLAUDE_CODE_SUBAGENT_EFFORT and its order, 2.1.316 reaching the stable channel, and November 30, when Opus 4.8 fast mode ends. The changelog is where each change shows up first.