SPM
Agents

The agent surface — appai + abc.

Every agent in the stack: the AppAI cockpit itself, the ABC skill family, per-codebase master agents, and the daily-workflow agents. Diagram is live, generated via /abc-diagrams.

Topology is hand-curated and lives at public/diagrams/agents-topology.html. Re-run /abc-diagrams to refresh.

Agent tiers

Four tiers, one stack.

Agents in the SPM stack split into four layers by scope. The higher tiers route to the lower ones — and every tier is just a skill at heart.

Master agents

Per-codebase super-agents. They own the project state, route work to sub-skills, and never lose context between sessions. /appai, /aibo, /aios, /villai.

/appai/aibo/aios/villai

Workflow agents

One agent per daily pipeline. Same shape every day: research → structure → render → log. /pad, /lad, /aad, /iad, /vad.

/pad/lad/aad/iad/vad

Specialist agents

Narrow scope, deep skill. /abc-architect, /abc-intake, /abc-proposal, /security-review, /persona, /skill-creator.

/abc-architect/security-review/persona

Org agents

Cross-codebase agents that speak for a whole org. cleverfox-agent, [client]-agent. They route to the right master agent based on the request.

cleverfox-agent[client]-agent
Agent anatomy

What every agent file declares.

The six fields you'll see in any well-shaped agent in the stack. Anything missing is a bug — explicit beats implicit when you're running fifty of them.

Trigger

When does this agent wake up? Cron, webhook, file change, PR event, Slack mention, or human invocation.

Memory

What does it remember between runs? File-based memory, vector store, or session-scoped only.

Tools

What can it touch? Read/edit/write, bash, web, MCP servers, specific APIs. Allow-listed per agent.

Persona

Who is it? Name, role, tone, refusal posture, escalation chain. Managed via PersonAI.

Escalation

What does it do when confidence drops? Pause, ask, fall back to a human, or refuse and log.

Budget

What's it allowed to spend? Per-call token cap, per-day cost cap, per-week kill switch.

Manage these with AgentAIPersona layer: PersonAI
Getting started

Your first agent in 5 steps.

Every agent in the stack started this way. The five steps below produce a working SKILL.md from scratch.

  1. 01

    Name the trigger

    When should it wake up? Daily cron, Slack mention, PR event, file change, or manual invocation. Ambiguous triggers produce ambiguous agents.

  2. 02

    Scope the memory

    What does it need to remember across runs? Keep it to one concern — project PBIs, content topics, or customer records. Never all three in one file.

  3. 03

    Allow-list the tools

    List every tool it needs: Read, Edit, Bash, WebSearch, MCP. Deny everything else by default. Start narrow and widen only when a real use-case demands it.

  4. 04

    Write the SKILL.md

    One file: trigger, memory schema, tool allow-list, persona (name + tone + refusal posture), escalation rule, and a worked example invocation.

  5. 05

    Run it, log it, tune it

    First run is the spec. Read the output against the intent. Fix the SKILL.md, not the output. Add a /review pass after 7 runs to close any drift.

Manage them all with AgentAI →

AppAI cockpit

· 2

MS26 PG Architect

Maintains 1G (Apps · Backoffice · Context) alignment across the fleet.

~/APPS/appai/.claude/agents/ms26-pg-architect.md

AppAI Cockpit

Build · preview · push · deploy orchestrator across all apps.

ABC skill family

· 19

ABC Master

Super master agent for the 1G ABC initiative.

~/.claude/skills/abc/

ABC Architect

Stack-aware system architect; produces ADRs and data models.

~/.claude/skills/abc-architect/

ABC Intake

Turns raw client input into a structured Project Card.

~/.claude/skills/abc-intake/

ABC Sector

Classifies projects into verticals and attaches sector packs.

~/.claude/skills/abc-sector/

ABC Proposal

Generates full proposal pack: markdown, HTML, JSX prototype, PDF, email.

~/.claude/skills/abc-proposal/

ABC Diagrams

Editorial-quality HTML+SVG diagrams, no Mermaid-slop.

~/.claude/skills/abc-diagrams/

ABC Social

Concept → 13-slide deck + AIBO page + video + email + LinkedIn + tweet.

~/.claude/skills/abc-social/

ABC SMM

URL → 6-asset social media content pack.

~/.claude/skills/abc-smm/

ABC JSX

Idea → single-file React JSX app shell artifact.

~/.claude/skills/abc-jsx/

ABC UXD

Canonical design tokens for all Mat Siems UI.

~/.claude/skills/abc-uxd/

ABC Tailwind

Tailwind v4 + shadcn rules for ABC apps.

~/.claude/skills/abc-tailwind/

ABC Shadcn

shadcn component management, docs, and registry.

~/.claude/skills/abc-shadcn/

ABC Next.js

Next.js 16 conventions — route groups, params as Promise, runtimes.

~/.claude/skills/abc-nextjs/

ABC Cleanup

Lint · unused-import sweep · file organisation.

~/.claude/skills/abc-cleanup/

ABC Env

Validate .env, compare environments, generate templates.

~/.claude/skills/abc-env/

ABC CCC

Claude Code Cheat — 10 POM crash course for teams.

~/.claude/skills/abc-ccc/

ABC GWS

Google Workspace cockpit with Fox-persona orchestration.

~/.claude/skills/abc-gws/

ABC MS-Context

Private 1G planning + world context pillar.

~/.claude/skills/abc-ms-context/

ABC UXD Audit

Full UX audit — mobile/tablet/desktop, a11y, consistency.

~/.claude/skills/abc-uxd-audit/

Codebase agents

· 6

AIBO

AIBO platform — personas, assets, media, workflows.

~/.claude/skills/aibo/

VillaI

V26 villa acquisition + ownership OS.

~/.claude/skills/villai/

AUDIO

Audio assets, DJ recordings, TTS, transcription, AAD.

~/.claude/skills/audio/

AIP

Onyx-based enterprise AI platform.

~/.claude/skills/aip/

CAC

Claude Architect Certification study coach.

~/.claude/skills/cac/

AMA

Anthropic Managed Agents study coach.

~/.claude/skills/ama/

Workflow agents

· 3

PAD

Page A Day — daily 5-stream context page.

~/.claude/skills/pad/

LAD

List A Day — fresh Top-100 list, never repeated.

~/.claude/skills/lad/

AAD

Audio A Day — narrated short with infographic.

~/.claude/skills/audio/
Common questions

Agents FAQ.

How is this different from running a single Claude chat?

A chat has no memory, no triggers, and no accountability. Agents in this stack each have a SKILL.md (their job description), a place they store output, and a defined invocation contract. That's the difference between a coworker and a stranger you have to brief every morning.

Do I need to write the agent myself?

No. We co-design the first three in the engagement — usually one writer, one researcher, one orchestrator. After that, you can fork the pattern or have us scaffold the next one. The fleet of 13 live sites ships entirely from this same template.

Where does the work actually run?

Locally for prototyping (Claude Code in your terminal), in production via the Claude API for managed agents, and on Vercel for any agent that needs a public surface. No proprietary runtime, no lock-in. If we vanish tomorrow, your agents keep working.

How are agents priced once they're live?

Managed Agents bill at £1 per agent-hour for compute + supervision. A typical site-ops agent uses 2-6 hours per month. Heavy media-generation agents (IAD, VAD, AAD) bill at usage cost + 20% — the rate sheet is on the pricing page.

What stops the agent going off-script?

Three things. (1) SKILL.md is opinionated — it says no as often as yes. (2) Every agent invocation logs to a JSON line we can replay. (3) Outputs land in a reviewed location (BO, GitHub PR, S3 bucket) — never directly to production. Drift is detected within one loop iteration.

Can I bring my own LLM?

Yes. The stack defaults to Claude (Opus 4.7 / Sonnet 4.6) but the skill format is provider-neutral. We've run the same agents against Gemini and GPT-5 for clients with policy constraints. Token economics are typically 20-40% better on Claude, but the architecture doesn't care.