Agents
The engine ships 54 agents in two fleets. The company fleet (27) runs the business: strategy, revenue, leads and funnel, quality, the production platform and the PWA lane, product references, product flow videography and media QA, and curriculum content. The stack fleet (27) writes the code: one agent per framework, layer, or tool.
Each fleet is a plugin, which is how it reaches other machines — see Onboarding.
| Fleet | Plugin | Agents | Lives in |
|---|---|---|---|
| Company | kun-company | 26 | .claude/agents/ |
| Stack | kun-stack | 27 | installed to ~/.claude/agents/ |
Company fleet (26)
The captain sits above everything and never writes code — it allocates, decides, and delegates.
| Agent | Scope |
|---|---|
| captain | CEO brain — weekly allocation, revenue strategy, coordination across products and the 7 humans |
Business (6)
| Agent | Scope |
|---|---|
| revenue | Pricing, proposals, contracts, cost analysis, revenue tracking |
| growth | Content strategy, SEO, social, developer relations, community |
| support | Customer onboarding, issue triage, knowledge base, SLA tracking |
| analyst | Market intelligence per vertical, competitors, usage analytics, churn |
| lead | Lead acquisition — contact-gap analysis, enrichment, tiering, outreach drafting against the CRM |
| funnel | Conversion — gate state, segmentation, value ladder, the stall clock from first reply to paying customer |
Product & delivery (6)
| Agent | Scope |
|---|---|
| product | Roadmap across the 5 products — stories, scope, prioritization, releases |
| tech-lead | Architecture across the repos — shared patterns, breaking changes, tech debt |
| ops | CI/CD, API spend, monitoring, infrastructure, uptime |
| guardian | OWASP Top 10, performance budgets, dependency health, SSPL compliance |
| cloudflare | The production platform — Workers, Containers, zone DNS, cron triggers, the balqalam.com lane |
| pwa | Installability, per-tenant manifests, service-worker caching that respects auth, offline outboxes, Web Push — see PWA |
Engine maintenance & media (8)
| Agent | Scope |
|---|---|
| quality | Routes the 17 niche quality keywords; owns /handover and /release |
| record | Screen-records demo videos and screenshot sets of real product flows, files into media archive |
| media-qa | Checks finished stills and videos before they ship — budgets from qa-pack, then Arabic shaping, RTL, brand, privacy and safe areas by eye; never edits pixels |
| report | Auto-fixes user-reported issues — read, verify, fix, close |
| analyze | Generates a tailored Claude Code config for any repo |
| learn | Extracts patterns, conventions, and team dynamics into memory |
| package | Cross-repo dependency audit and upgrade coordination |
| clone | Translates a captured live-URL section into house-stack JSX |
Content (2)
The textbook lane — see Textbook. Narrow toolsets by design: Read, Write, Glob — a reader that cannot run a shell cannot peek.
| Agent | Scope |
|---|---|
| transcribe | Reads scanned page images and writes one Markdown file per page, exactly as printed — never corrects, never guesses |
| adjudicate | Settles every disagreement between two reads against the page and its crops, repairs the page, records the verdict — the only agent that edits a transcribed page |
Product references (4)
Read-only pattern sources — ask them how an existing product solved something.
| Agent | Product |
|---|---|
| hogwarts | Education SaaS — multi-tenant, LMS, SIS, billing |
| souq | E-commerce — multi-vendor marketplace, cart, vendor dashboards |
| mkan | Rentals — property listings, booking, search |
| shifa | Medical — appointments, patient records, healthcare workflows |
Stack fleet (27)
| Group | Agents |
|---|---|
| Framework | nextjs · react · typescript · tailwind · prisma · authjs · shadcn |
| Design | architecture · orchestration · pattern · structure |
| UI | atom · template · block · icon · semantic |
| DevOps | build · deploy · test · git · github |
| Platform | middleware · internationalization · sse · optimize · performance · comment |
How an agent is defined
An agent is a markdown file with YAML frontmatter:
---
name: prisma
description: Prisma 6 expert — PostgreSQL ORM, migrations, query optimization
model: sonnet
effort: medium
---
Instructions, context, and domain knowledge...model is always an alias (opus / sonnet / haiku), never a pinned version — the alias resolves through model_tiers in .claude/engine.json, so upgrading the fleet is a one-line change. effort sets the reasoning budget per delegation.efforts.
| Tier | Used for |
|---|---|
| opus | Architecture, orchestration, review, strategy, leadership |
| sonnet | Standard build agents with known patterns |
| haiku | Mechanical work — comment, git, icon, github, semantic, sse, structure |
The main loop itself runs Opus 5.5 (Fable 5.1 only as a /model escalation); agents stay on the aliases above, which resolve forward on their own.
Routing
You don't name agents. You say a keyword and the engine routes:
"saas billing"
→ orchestration → prisma (schema) + block (UI) + stripe MCP
"ship hogwarts admission"
→ captain → product (stories) → tech-lead (architecture)
→ orchestration → stack specialists
See Keywords for the full registry.
Installation
cd ~/kun && bash .claude/scripts/setup.sh engineerEvery machine gets both fleets. Role is a label and a secrets-trust tier, not a capability gate.