Agent Reference

Agent reference

The seats this harness can install. A run installs between 7 and 15 of them, decided by the contract and by what the codebase analysis found, plus one dev-agent per module that actually exists. Never all of them by default: see Tailored Build.

Every seat carries an explicit model and effort. An unset model means inherit, which silently bills mechanical work at the caller's tier, so the harness never leaves it unset.

Seat	Model	Effort	maxTurns	Installed	What it is for
ba-analyst	sonnet	high	40	always	Writes and edits the specification set in docs/specs/ and the PRDs in docs/requirements/.
brainstormer	sonnet	high	20	needs long	Structured brainstorming for an open business or technical decision - frames it, diverges to options, scores a trade-off matrix, recommends.
code-reviewer	opus	high	25	absent when solo_review	Reviews a branch diff against coding standards and project rules.
data-modeler	sonnet	high	40	needs db	Designs and modifies the the ORM schema and its migrations, per the ERD and the data dictionary.
db-engineer	sonnet	medium	40	needs db and db_engineer	Database operations - applies and troubleshoots migrations, tunes queries and indexes, checks integrity, manages the local DB environment.
db-seeder	haiku	low	15	needs db and db_seeder	Seeds the local or dev database with deterministic synthetic data.
debugger	opus	xhigh	30	always	Root-cause diagnosis of failing CI jobs, failing tests, runtime errors, and environment problems.
devops	sonnet	medium	40	always	CI/CD (CI), environments, infrastructure, and gated releases to the hosting target.
history-tracker	haiku	low	10	needs audit or long	Inspects and summarizes the agent-run archive in .claude/state/history/ (written automatically by the agent-history hook).
merge-manager	opus	high	20	always	Merges approved pull requests into the default branch and resolves merge conflicts under delegated authority.
orchestrator	opus	high	60	always	Mission controller for cross-domain work.
qa-test	sonnet	medium	40	needs tests	Writes and runs the automated tests (unit: the unit framework, e2e: the e2e framework) that prove the stated acceptance criteria.
reviewer	opus	high	25	needs solo_review	The merged review gate for small projects - one read-only pass over a branch diff covering both code quality and security.
security-reviewer	opus	high	25	absent when solo_review	Reviews a branch diff for security and privacy - secrets, PII, authn/authz, input validation, injection, dependency risk.
spec-guardian	sonnet	medium	15	always	Checks a feature or change against the written requirements - FR/NFR, acceptance criteria, business rules.
tech-researcher	sonnet	medium	20	needs long	Technology research and evaluation with cited evidence - libraries, providers, approaches - against this project's constraints.

Reviewers hold no write tools

code-reviewer and security-reviewer (or the merged reviewer under the solo_review flag) have no Edit and no Write. Not "usually" - none. A reviewer that can edit is not a review gate.

The template seat

dev-agent.md is not in the table because it is a template, instantiated once per module with that module's real paths as its scope. A module with no seat, or a seat with no module, is a finding in harness-view assess.