Agents Reference

ulk ships 90 specialized agents organized into 12 categories and 7 phases. Each agent is a markdown file in framework/agents/ with a YAML frontmatter specifying its name, model, phase, tools, and invocation protocol.

Overview

Dimension	Count
Total agents	90
Categories	12
Phases	7
Agents with persistent memory	9
Agents available as Managed Agents (cloud)	5

Category	Agent count	Purpose
orchestrators	6	Entry points, routing, multi-agent coordination
docs	6	Spec, todo, README, memory, sync
audit	17	Code quality, security, accessibility, cost, UX
session	11	Checkpoints, diagnostics, error fixing, task execution
mobile	3	iOS, macOS, Android
sync	2	Notion, Linear
analyze	6	Stack-specific analysis (Next, Nuxt, Astro, SPIP, SwiftUI)
deploy	6	Vercel, Netlify, Cloudflare, AWS, Docker, Fly
routines	2	Cloud scheduling, CI guard
test	2	Unit tests, E2E tests
frontend	10	UI generation, design system, QA, migration
vps	17	Full VPS lifecycle management

Phases

Phase	Description	Example agents
`define`	Project setup, brief, stack decisions	Tony, Shuri, Stark
`plan`	Task decomposition, roadmap	Task Runner, Routine
`build`	Implementation, code generation	Brique, Fluke, Steve
`verify`	Testing, spec conformance	CI Guard, Verify, frontend-qa
`review`	Audits, quality checks	Sargeras, ED-209, Frodo
`ship`	Deploy, commit, release	2b3, Killbill, deploy-*
`orchestrator`	Multi-agent coordination	Bruce, Blackemperor, Lovecraft

Invocation Syntax

All agents are invoked as slash commands in Claude Code:

/ulk:<agent-name>

Examples:

/ulk:bruce
/ulk:bruce go
/ulk:bruce next
/ulk:sargeras
/ulk:2b3
/ulk:godspeed

Some agents also respond to natural language aliases. Bruce (25) accepts commands like go, status, next, fix, audit, ship, and sync. Others accept mode flags: strange mode=prompt, lovecraft memory surface, killbill mode=detach-github.

Orchestrators

Bruce (25) — Primary orchestrator

The single entry point for 90% of use cases. Bruce understands the project state, asks the right clarifying questions, and routes to the appropriate agents without the user needing to know which agent handles what.

/ulk:bruce go        # Detect project state and take the most logical action
/ulk:bruce status    # Full project diagnostic
/ulk:bruce next      # Pick and execute the next task from todo.md
/ulk:bruce fix       # Find and fix errors (delegates to Robocop)
/ulk:bruce audit     # Code quality audit
/ulk:bruce audit all # 4 parallel audits: code, perf, a11y, SEO
/ulk:bruce review    # Pre-delivery completeness review
/ulk:bruce ship      # Pre-release checks
/ulk:bruce sync      # Sync with Notion/Linear

Bruce has persistent memory between sessions. It remembers the project, stack, and preferences. Model: Opus, effort: xhigh.

Banner Mode: Bruce activates a higher-scrutiny mode on critical signals — security findings, scope creep, context above 80%, audit warnings. In Banner Mode it triple-checks every decision (why / risks / downstream impact) before routing.

Blackemperor (18) — Multi-mode parallel orchestrator

Runs multiple agents simultaneously and produces a consolidated report. Supports multiple modes:

/ulk:blackemperor          # Default: audit mode (4 agents in parallel)
/ulk:blackemperor          # Also handles: build, deploy, review modes

Model: Opus, effort: xhigh.

Lovecraft (47) — Memory orchestrator

Manages the cross-session knowledge vault (docs/_memory/).

/ulk:lovecraft memory capture    # Capture MEMORY.md → docs/_memory/<category>/
/ulk:lovecraft memory surface    # Read-only summary for current session
/ulk:lovecraft memory distribute # Inject vault into CLAUDE.md block
/ulk:lovecraft memory dream      # REM-style consolidation: dedup, normalize, archive

Model: Sonnet.

Tony (50) — Engineer-in-chief

Brief → stack recommendation + architecture + timing estimate. Invoked by Bruce when a new project or major feature needs architectural decisions.

/ulk:tony

Model: Opus, effort: xhigh. Has persistent memory for stack preferences and project history.

Mathieu (61) — Senior Product Strategist

Product/business/UX strategy, LLM integration audit, roadmap advising. Complementary to Tony (architecture) and Stark (design).

/ulk:mathieu

Model: Opus, effort: xhigh.

Project Decomposer (35)

Transforms a todo/spec into self-contained Claude Code prompts. Each prompt works without prior context — a defense against context drift.

/ulk:project-decomposer

Audit Agents

Sargeras (45) — 10-axis omniscient audit

Full project audit across 10 axes: security, performance, architecture, tests, CI/CD, accessibility, SEO, documentation, tech debt, cost. Produces a scored report with violations sorted by severity.

/ulk:sargeras

Model: Opus, effort: xhigh. Available as a Managed Agent (cloud, weekly cron).

ED-209 (52) — Security audit

OWASP Top 10 focused security audit: exposed secrets, vulnerable dependencies, attack surface, HTTP headers, CORS, authentication.

/ulk:ed209

Model: Opus, effort: xhigh. Available as a Managed Agent (cloud, monthly cron).

Frodo (62) — Generational UX audit

Evaluates UI across 5 generational cohorts (Boomers → Gen Alpha) on 5 UX dimensions. Produces a canonical generational audit report in docs/generational-audit/.

/ulk:frodo

Model: Opus, effort: xhigh.

Harper (66) — Adversarial review

Systematic devil’s advocate review. Challenges assumptions, surfaces blind spots, provides structured counter-arguments.

/ulk:harper

Model: Sonnet.

Agathe (60) — Design guardian

Art director responsible for the design system and layout quality. Reviews and advises on any design decision. Orchestrates design agents (Stark, Agamotto, Brique, Visual-Auditor) around docs/design.md.

/ulk:agathe

Model: Opus, effort: xhigh.

Killbill (56) — Cost killer

Audits Vercel, GitHub, and Neon costs. Identifies waste, produces a savings plan. Has a killswitch mode that can actually detach GitHub Actions, Dependabot, scanning, and webhooks via gh api.

/ulk:killbill
/ulk:killbill mode=detach-github

Model: Opus, effort: xhigh.

Other audit agents

Agent	Role
a11y-auditor	WCAG 2.1/2.2 accessibility audit
perf-auditor	Core Web Vitals, bundle size, database performance
seo-auditor	SEO technical audit
vision	Code quality across 8 dimensions
alex (59)	Music-tech advisor (Laravel + AI + music APIs)
benjamin (64)	Devil’s advocate and due-diligence lens
minitel (67)	UX writing / microcopy audit (voice & tone)
context-audit	Token health score 0-100, detects waste
cutkiller (51)	GitHub detach operations audit
picsou	Cost planning and budget forecasting
astride	Project definition review
amiral	Strategic review
code-graph	Code review graph analysis
tools-checker	CLI and skill diagnostic
claude-md-optimizer	CLAUDE.md quality and token efficiency

Session Agents

Godspeed (00) — Diagnostic

Scans the project in seconds. Caches the result for 30 minutes. Used by Bruce at startup. Model: Haiku, effort: low (read-only, structured output).

/ulk:godspeed

2b3 (08) — Session checkpoint

End-of-session pipeline: lint → tests → simplification → memory capture → commit. Integrates Verify (65) as a pre-commit gate.

/ulk:2b3

Model: Sonnet. Available as a Managed Agent (daily cron at 18:00).

Task Runner (04) — Task executor

Reads tasks from docs/todo.md, executes them one at a time, tracks velocity per category, refines estimates.

/ulk:task-runner

Model: Sonnet.

Robocop (11) — Error fixer

Invoked on broken builds or red tests. Identifies the root cause, applies fixes, re-runs checks.

/ulk:robocop

Model: Opus, effort: xhigh.

CI Guard (54) — CI/CD auto-fix

Monitors CI failures and attempts automatic fixes. Available as a Managed Agent triggered by check_suite.completed (failed) GitHub webhook.

/ulk:ci-guard

Model: Sonnet.

Gandalf (34) — Context health monitor

Checks compliance with the 5 context hygiene rules. Alerts when context exceeds 40%.

/ulk:gandalf
/ulk:gandalf status    # Quick assessment
/ulk:gandalf save      # Guide to persist state before /clear
/ulk:gandalf rules     # Recall the 5 hygiene rules

Model: Haiku, effort: low.

Xavier (57) — Context verifier

Verifies the work context: correct repo, correct account, last git state, last session state. Can run automatically at session start via the --with-xavier-hook hook.

/ulk:xavier

Model: Sonnet.

Verify (65) — Spec-to-code conformance

Answers one question: does the delivered code match what was specified? Operates on a 3-dimension × 3-severity matrix (Completeness × Correctness × Coherence, severities: CRITICAL / WARNING / SUGGESTION).

/ulk:verify

Wired automatically into 2b3 (Phase 4.7), Bruce (Phase 5.0 pre-audit gate), and Task Runner (Phase 3.1.5 pre-done gate). Exit codes: 0 (clean), 1 (CRITICAL — blocks archive), 2 (card not found).

Other session agents

Agent	Role
rodin (46)	Socratic questioning to clarify requirements
sensei	Knowledge transfer and onboarding
jean-claude (63)	Virtual support agent — triage GitHub issues and Linear tickets
context-audit	Session token health score and waste detection

Doc Agents

Shuri (01) — Documentation pipeline

Generates and maintains docs/spec.md, docs/todo.md, and syncs CLAUDE.md / README.md. Primary doc orchestrator.

/ulk:shuri
/ulk:shuri mode=full     # Full pipeline including 2b3 + /clear suggestion
/ulk:shuri mode=refactor # Migrate from legacy obsidian to faru doc-mode

Model: Sonnet.

Strange (16) — Reverse documentation

Reconstructs docs/spec.md from existing code when documentation is missing or outdated. Also handles prompt reverse-engineering: given an AI output, infer the prompt that produced it.

/ulk:strange
/ulk:strange mode=prompt

Model: Opus, effort: xhigh.

Other doc agents

Agent	Role
agamotto (17)	Reverse-engineers design from Figma URLs into docs/design.md
friday	Session snapshots
obsidian-vault (39)	Obsidian vault management
marketing-maestro	Marketing copy and content generation

Frontend Agents

Agent	Phase	Role
stark (58)	define	Design system generation from brand/URL/screenshots
brique	build	Component implementation guided by docs/design.md
frontend-orchestrateur	orchestrator	Frontend sub-task routing
frontend-qa (02)	verify/review	Functional, performance, visual testing
visual-auditor (03)	verify	Visual regression and screenshot audits
backoffice-auditor	review	Back-office UI audit
pencil-generator	build	Wireframe generation
ranma	plan	Frontend task planning
shadcn-migrator	review	shadcn/ui migration audit
svg-analyzer	build	SVG optimization

Specialist Agents

Mobile (3 agents)

Agent	Role
happy	Mobile project diagnostic and routing
steve (27)	iOS/macOS development (5 Apple platforms: macOS, iOS, iPadOS, watchOS, tvOS)
fluke (48)	Android development (Material 3, Jetpack Compose)

Sync (2 agents)

Agent	Role
bifrost	Notion sync
brigitte	Linear sync

Deploy (6 agents)

Each deploy agent handles one target:

/ulk:deploy-vercel
/ulk:deploy-netlify
/ulk:deploy-cloudflare
/ulk:deploy-aws
/ulk:deploy-docker
/ulk:deploy-fly

Analyze (6 agents)

Stack-specific analysis agents:

/ulk:analyze-next     # Next.js
/ulk:analyze-nuxt     # Nuxt.js
/ulk:analyze-astro    # Astro
/ulk:analyze-spip     # SPIP
/ulk:analyze-swiftui  # SwiftUI
/ulk:benjamin         # Devil's advocate analysis

Test (2 agents)

/ulk:test-unit   # Unit test generation and execution
/ulk:test-e2e    # End-to-end test generation and execution

VPS (17 agents)

A complete VPS lifecycle suite managed by orchestrateur-vps. Individual agents cover: audit, backups, CI/CD, cleanup, compliance, cost, deployment, Docker, documentation, environments, incidents, installer, migration, monitoring, network, security.

/ulk:orchestrateur-vps    # Routes to the right VPS agent

Agent Models and Effort

ulk enforces a model and effort policy at the frontmatter level. The key rule: do not set the CLAUDE_CODE_SUBAGENT_MODEL environment variable — it would override all agents’ model: declarations.

Tier	Model	Effort	Agents
Heavy orchestrators and deep audits	Opus	xhigh	bruce, blackemperor, lovecraft, tony, mathieu, sargeras, ed209, agathe, frodo, killbill, strange, agamotto, happy, steve, fluke, stark, robocop, rodin, benjamin, frontend-orchestrateur
Lightweight scanners	Haiku	low	godspeed, gandalf, tools-checker
Everything else	Sonnet	high (default)	64 agents

For interactive sessions, /model opus is recommended (Bruce orchestrates better outside plan mode).

Persistent Memory (9 agents)

These agents retain state between sessions via ~/.claude/agent-memory-local/<agent>/:

Agent	What it remembers
Bruce (25)	Project, stack, preferences
Godspeed (00)	Last diagnostic (expires at next commit)
Gandalf (34)	Context zone, alerts
Task Runner (04)	Velocity per category
2b3 (08)	Recurring problems
Steve (27)	Mobile API decisions
Fluke (48)	Android stack, form factors
Tony (50)	Preferred stacks, project history
Lovecraft (47)	Full vault in docs/_memory/

Agents Reference

Overview

Categories

Phases

Invocation Syntax

Orchestrators

Bruce (25) — Primary orchestrator

Blackemperor (18) — Multi-mode parallel orchestrator

Lovecraft (47) — Memory orchestrator

Tony (50) — Engineer-in-chief

Mathieu (61) — Senior Product Strategist

Project Decomposer (35)

Audit Agents

Sargeras (45) — 10-axis omniscient audit

ED-209 (52) — Security audit

Frodo (62) — Generational UX audit

Harper (66) — Adversarial review

Agathe (60) — Design guardian

Killbill (56) — Cost killer

Other audit agents

Session Agents

Godspeed (00) — Diagnostic

2b3 (08) — Session checkpoint

Task Runner (04) — Task executor

Robocop (11) — Error fixer

CI Guard (54) — CI/CD auto-fix

Gandalf (34) — Context health monitor

Xavier (57) — Context verifier

Verify (65) — Spec-to-code conformance

Other session agents

Doc Agents

Shuri (01) — Documentation pipeline

Strange (16) — Reverse documentation

Other doc agents

Frontend Agents

Specialist Agents

Mobile (3 agents)

Sync (2 agents)

Deploy (6 agents)

Analyze (6 agents)

Test (2 agents)

VPS (17 agents)

Agent Models and Effort

Persistent Memory (9 agents)