ReleaseLatest release v0.10.2 · 24 d’oct. del 2026These pages describe v0.10.2, the published release.Changelog →

Run agents in parallel

Codewhale can hand a focused piece of work to a sub-agent and keep going while it runs. Use this when a task splits into independent parts — mapping a codebase, reviewing a change, running tests — so they happen at the same time.

Ask for it

In Work, ask Codewhale in plain words to delegate independent parts of a task. Say what you want split up:

Prompt

Use three explore agents in parallel: one maps the auth code,
one maps the billing code, one lists the tests that cover both.
Then summarize what would break if we changed the session format.

Each sub-agent starts in the background and reports back when it finishes. Your composer stays free, and the parent turn picks up the results.

Pick a role

A role is the stance a sub-agent takes. Name one in your request, or let Codewhale choose. A sub-agent never gets more permission than your session has.

general
Does whatever the brief says; may edit and run commands. The default.
explore
Read-only. Maps the relevant code fast — “find every caller of this function.”
planner
Designs an approach without changing anything.
reviewer
Reads and grades a change, with a severity for each finding.
implement
Lands one specific change with the smallest edit.
test
Runs tests and checks, then reports pass or fail. Does not edit code.
advisor
Short, careful second opinion on a judgment call. No commands.
custom
Only the tools you list, for tightly limited jobs.

To give a role a specific model every time, save it with /fleet setup — see Run a workflow.

Keep parallel edits apart

When two sub-agents would edit the same repository, ask for each to work in its own worktree. Codewhale creates a fresh git worktree and branch for that agent beside your repository, under .codewhale-worktrees/, so your checkout stays clean until you merge.

A worktree is isolation, not permission: a writer still needs a writing role and the paths it may change. When a new write claim overlaps an active claim in the same checkout, the host refuses the new claimant; it does not rewind or pause the existing writer.

Watch and answer them

Codewhale

/subagents

/subagents (the same view as /fleet workers) lists the agents attached to this session and what each one is doing. Select one to read its transcript.

Sub-agents follow your approval setting and may be more restricted. In Ask, a call that needs approval appears in your session while the agent waits. Auto-Review handles tool approvals through policy and eligible independent review; blocked calls or failed reviews are refused. Check the child transcript for decisions and refusal reasons. OS and connected-service permissions may still need your action.

Know the limits

Running at once
64 by default. Set max_subagents (up to 128) in ~/.codewhale/config.toml.
Queued plus running
Up to 1,024 admitted by default, configurable with [subagents] max_admitted. Admitted agents wait for an execution slot; launches beyond the admission limit are refused.
Nesting
No default product depth cap. Set [subagents] max_depth to an owner-selected limit; child requests cannot widen an inherited restriction.

These are capacity controls, not targets. Keep each agent's scope clear. Durable Fleet runs record work across manager exits, but recovery is explicit: reconcile stale leases, inspect the result, then choose whether to restart a worker. They do not keep executing while your machine is asleep.

Next

Run a workflow

Turn a plan you repeat into a Workflow with phases and a record of each run.

Track progress

How the To-do list shows what is done, in progress, and left.

Review what changed

Check what the agents changed before you keep it.