Investigator.
Finds the files, patterns, ownership boundaries, and risks in an unfamiliar codebase.
Subagents
Many minds. One outcome.
Specialists for the parts of software work that should not wait in line. Focused brushes. One canvas.
reviewer
test-writer
▾migrator
▾Map every export call site
Write a failing test for the timeout
▾Add retry with backoff
▾Run the full suite
∠maintools4ctx28k / 1Mcache0 cachedmodeautoworking
Specialists, built in.
Reviewer.
Looks for bugs, regressions, incomplete behavior, and missing tests before the main session ships.
Security.
Checks data flow, permissions, secrets, and unsafe execution edges.
Docs.
Turns implementation detail into useful explanation for maintainers and teammates.
Tests.
Adds focused coverage around changed behavior and verifies the result.
Migration.
Breaks large moves into stages the main canvas can inspect and verify.
Browser.
Looks at the UI the way a user would, then reports concrete visual findings.
Generalist.
Handles bounded work that needs no specialist role.
Fan out. Decide once.
Picasso can split independent work, wait for results, then bring everything back into one decision. Subagents are useful when tasks can run in parallel without blocking the next local step, and every result is tied back to the parent session for replay.
developer flowterminal to Studio
repoopen canvas
sketchreview plan
checktests pass
replayinspect later
Fork vs named.
| Mode | Context | Prompt | Use |
|---|---|---|---|
| Fork. | Inherits parent context. | Same role. | Continue a known thread. |
| Named. | Fresh context. | Role-specific. | Focused parallel work. |
Define your own brushes.
Teams can create project-specific roles with scoped tools, model choices, memory access, step budgets, and permission boundaries. The role stays explicit so the result is easy to review.
name: dependency_auditormodel: sonnettools: [read_file, grep, shell]max_steps: 40Available from CLI, TUI, and Mac.
Slash commands, runtime goals, and the Mac app all use the same subagent catalog and session bridge. A subagent launched from the TUI or Mac app records the same parent goal, evidence, permissions, and result shape.
Named.
Pick a catalog role for review, security, tests, docs, browser, or migration.
Fork.
Clone the parent context to try another path without losing the original.
Scoped.
Limit tools, files, models, memory, and steps per role.
Subagents ask the way you do.
Work handed to a subagent follows your approval settings on the CLI and the Mac. If a delegated step needs your say-so, it reaches you instead of being refused out of sight.
See the team on the Mac.
When agents work as a team, the inspector lists the roster and a task board layered by dependency, with each task's owner and whether it is ready, in progress, or waiting on others. Terminals the agent opens are listed read-only beside it.
Roster.
Who is on the team and what each member is doing.
Task board.
Tasks ordered by what depends on what, with owners.
Agent terminals.
Every shell the agent started, with its state and folder.