Skip to content

Agents reference

The skill ships 29 agent definitions in agents/: 27 dispatchable subagents plus 2 shared "core" preambles that are prepended to other agents and never dispatched directly. Your host agent (Claude Code, OpenCode, …) picks them up automatically after memory-bank install.

All agents follow an evidence-before-claims discipline: no "tests pass" without the actual test output, no "done" without verification.

Dev-role agents (used by /mb work)

/mb work matches the stage description against a role and dispatches the matching specialist. Every dev-role agent receives the mb-engineering-core discipline preamble (TDD, Clean Architecture/FSD, SOLID, no placeholders) and the mb-tooling-core routing preamble (graph-first code navigation).

Agent Specialty When it's picked
mb-developer Generic implementer Default when no specialist matches
mb-backend APIs, services, DB, async/concurrency Backend-shaped stages
mb-frontend React/Vue/Svelte/Solid, a11y, responsive UI Frontend-shaped stages
mb-ios SwiftUI/UIKit, Combine, async/await iOS stages
mb-android Jetpack Compose, coroutines, Hilt, Room Android stages
mb-devops CI/CD, Docker, K8s, Terraform, observability Infra stages
mb-qa Test design, coverage strategy, edge cases Test-focused stages
mb-analyst SQL, dashboards, ETL, A/B analysis Data/metrics stages (no production code)
mb-architect Domain modelling, ADRs, refactoring strategy Design stages (doesn't ship features alone)

Verification & quality gates

Agent Role Invoked by
plan-verifier Audits the diff against the plan's DoD, item by item /mb verify (required before /mb done for plan work)
mb-test-runner Runs tests, parses output into strict JSON; never reports "not-run" as pass /test, plan-verifier
mb-rules-enforcer Deterministic SRP / Clean Architecture / TDD-delta checks on changed files /review, /commit, /pr, plan-verifier
mb-judge Independent final gate: GO / GO_WITH_BACKLOG / NO_GO /mb work --judge and governed workflows

Review ensemble (opt-in via --review or governed workflows)

Agent Focus
mb-reviewer Single-reviewer mode: structured verdict (APPROVED / CHANGES_REQUESTED) with severity-classified issues
mb-reviewer-logic Requirement coverage, behavior, edge cases, regressions
mb-reviewer-quality Maintainability, SOLID/DRY/KISS/YAGNI, architecture boundaries
mb-reviewer-security Secrets, injection, authz/authn, filesystem/network risks
mb-reviewer-tests TDD evidence, regression coverage, test determinism
mb-reviewer-scalability Complexity, hot paths, memory, concurrency, IO
mb-reviewer-lead Synthesizes the aspect reports into one canonical review

Memory & knowledge agents

Agent Role Invoked by
mb-manager Maintains .memory-bank/ core files (actualize, search, notes, tasks) /mb context\|search\|note\|tasks\|update\|done, PreCompact hook
mb-doctor Finds & fixes inconsistencies across core files /mb doctor
mb-codebase-mapper Writes codebase/*.md docs (stack/arch/quality/concerns) /mb map
mb-wiki-author Writes one wiki article per code community (Haiku-tier) /mb wiki
mb-wiki-synthesizer Finds surprising cross-community connections (Sonnet-tier) /mb wiki

Research agents (never author plans or code)

Agent Role
mb-research Codebase + web research router: graph for structure, semantic search for concepts, /mb recall for decisions, web for the rest
mb-researcher Ecosystem research, option matrices, technical due diligence before planning

Shared cores (not dispatchable)

File Purpose
mb-engineering-core Engineering-discipline preamble prepended to every dev-role agent
mb-tooling-core Graph-first code-navigation routing prepended to every dev-role agent

See also