Roles
9 roles. A role is a bounded kind of worker; its tier says how much model capacity it gets, and .act/tiers.json maps tiers to concrete models only at generation time.
builder
Section titled “builder”Implements a bounded assignment - code, migration, tests, configuration - and returns a result plus evidence; never commits.
- Tier:
standard - Reasoning:
medium - Tools:
Read, Write, Edit, Bash, Grep, Glob
Implements a bounded assignment: code, migration, tests, configuration. Applies R-role-worker.
Source: .act/agents/builder.md
debugger
Section titled “debugger”Searches for a bug’s cause by hypothesis rather than guesswork; reproduces first, separates symptom from cause.
- Tier:
standard - Reasoning:
high - Tools:
Read, Bash, Grep, Glob
Searches for the cause of a reported bug that the orchestrator describes. Fixes nothing — the fix
is a separate assignment (builder). Applies R-role-worker.
Source: .act/agents/debugger.md
doc-writer
Section titled “doc-writer”Maintains docs/project/ (never docs/ai/) - works findings into the project docs, keeps cross-references and status markers current.
- Tier:
standard - Reasoning:
medium - Tools:
Read, Write, Edit, Grep, Glob, Bash
Maintains docs/project/ — project documentation, not the collaboration workspace under
docs/ai/. Applies R-role-worker.
Source: .act/agents/doc-writer.md
expert-solver
Section titled “expert-solver”High-reasoning escalation, called only after a worker has failed the same task twice or hit an unsolvable error.
- Tier:
expert - Reasoning:
max - Tools:
Read, Edit, Write, Bash, Grep, Glob
Escalation only, per R-role-escalate — called after a worker has failed the same task twice, or
an edge case has a standard worker stuck. Senior architect and problem-solver for exactly that
case, not routine implementation. Applies R-role-worker.
Source: .act/agents/expert-solver.md
explorer
Section titled “explorer”Read-only codebase research across multiple files and directories; reports findings backed by path:line.
- Tier:
standard - Reasoning:
low - Tools:
Read, Grep, Glob, Bash
Read-only research across multiple files and directories; findings backed by <path>:<line>.
Applies R-role-worker.
Source: .act/agents/explorer.md
optimizer
Section titled “optimizer”Polishes freshly written code for brevity and readability - at most two rounds, no algorithm tuning.
- Tier:
standard - Reasoning:
medium - Tools:
Read, Edit, Bash, Grep, Glob
Runs after builder, only on the code that assignment just wrote — files and lines named in the
assignment, never grown code from elsewhere and never project-wide. Applies R-role-worker.
Source: .act/agents/optimizer.md
quick-check
Section titled “quick-check”Fixed, read-only lookups without judgment (git status, tests, files, line counts).
- Tier:
light - Reasoning:
none - Tools:
Read, Grep, Glob, Bash
Runs a fixed set of read-only lookups and returns the raw result, without judgment. Applies
R-role-worker.
Source: .act/agents/quick-check.md
reviewer
Section titled “reviewer”Adversarial review before a commit — bugs, style, and task fidelity — plus ALLOW/BLOCK on a flagged safeguard call.
- Tier:
elevated - Reasoning:
high - Tools:
Read, Bash, Grep, Glob
Adversarial review before a commit, plus ALLOW/BLOCK on a flagged tool call. Applies
R-role-worker.
Source: .act/agents/reviewer.md
test-writer
Section titled “test-writer”Writes tests to existing code, or test-first from a concept/interface alone; checks behavior, not implementation.
- Tier:
standard - Reasoning:
medium - Tools:
Read, Edit, Write, Bash, Grep, Glob
Writes tests — to existing code, or test-first from a concept/interface description alone — and
proves them with a test run. Applies R-role-worker.
Source: .act/agents/test-writer.md
Tier mapping (Claude Code)
Section titled “Tier mapping (Claude Code)”| Tier | Model alias |
|---|---|
light |
haiku |
standard |
sonnet |
elevated |
opus |
high |
opus (reasoning one step higher) |
expert |
fable |