Home / Several agents and adversarial checking

Isolating and running in the background: worktrees and the dry loop

A git worktree per agent stops the files of several parallel agents from colliding, a dry loop repeats until a cycle no longer returns anything new, judged against an already seen set, and since 13 August 2026 a non teammate agent launched during an interactive session runs in the background by default with a notification at the end rather than blocking the session.

An agent that writes into the same repository as another agent, at the same time, risks overwriting what the other has just produced. The fix documented by Claude Code is to give each parallel agent its own temporary copy of the repository, a git worktree, by adding isolation: worktree to its configuration header. Each agent then works in its own copy, and this copy cleans itself up automatically when the agent has made no changes.

A loop that knows when to stop

A second move answers a different problem, a search that does not know when it has finished finding things. The official documentation describes this pattern for spotting flaky tests: rerun the suite several times, note which tests fail intermittently, and stop as soon as two consecutive runs no longer flag anything new. The same principle applies to any exhaustive search, a dependency audit, an inventory of orphan files: each cycle compares its result against the set already known, and only an item missing from that set counts as new for that round.

Cycle 1 : 4 elements trouves, 4 nouveaux, ensemble connu = 4
Cycle 2 : 5 elements trouves, 1 nouveau, ensemble connu = 5
Cycle 3 : 5 elements trouves, 0 nouveau
Cycle 4 : 5 elements trouves, 0 nouveau, deux tours sans nouveaute, la boucle s'arrete

The change of 13 August 2026

The switch to background running is not a setting you turn on, it has become the default behaviour. Since Claude Code version 2.1.232, released on 13 August 2026, an agent that is not a teammate of the session and starts during an interactive session runs in the background instead of blocking the ongoing conversation; the changelog does not say whether it gives a notification at the end, that point still needs checking independently. The same changelog places this change in the same entry as the default activation of the mode that makes a subagent inherit the cache and the full conversation, joined by a plain and: it establishes no causal link between the two, these are two simultaneous changes in the same version, not one mechanism triggering the other.

Worktree isolation and background execution combine naturally: several agents launched in parallel, each in its own copy of the repository, now run without blocking the session while they work. The previous lesson on choosing the architecture determines whether these agents should be launched in parallel, this one determines how to make them coexist without colliding. The next lesson deals with what happens once their number goes beyond a handful.

One detail changes how reliable the dry loop is: the count of new items is calculated against the set accumulated since the very first cycle, never against just the previous cycle. An item that disappears then reappears in the following cycle must not be recounted as new, or the loop risks running indefinitely on an unstable result rather than on a search that is genuinely incomplete.