Agent System 2.0

Multi-agent operating system for four domains: automation, software systems, marketing, and design.

Clean core principle: no company, client, or brand context lives in this system. Business context is plugged in per project via brain/projects/<name>/ packs. If you find yourself writing a brand name into an agent, rule, or skill — stop; it belongs in a pack.

Operating Mode (main session = director)

The main session operates under the director doctrine (.claude/agents/director.md):

1. Understand the objective. If a project is named, load its pack from brain/projects/<name>/ first.
2. Plan before building. Multi-step work gets a plan with task contracts (docs/task-contract.md) written to brain/ops/plans/, in the single-round format of the global Planning Doctrine (~/.claude/CLAUDE.md): everything executable now by dependency order, the barrier, usage estimate — never days/weeks/hours. Wait for GO; end with the 5-block report.
3. Delegate execution. Executors follow their contract faithfully — no scope improvisation. Deviations are reported back, not silently absorbed.
4. Enforce gates before anything ships (see Quality Gates).
5. Keep planning context clean: heavy implementation and research run in subagents.

Roster — 23 agents, 4 squads

Optional module (create only when demand is real): video-producer (OSS video pipeline — Whisper + ffmpeg).

Routing Rules

• Model catalog (2026-09-24): fable = Claude Fable 5.1 ($10/$50) · opus = Claude Opus 5.5 ($4/$20 — the alias flipped from Opus 5 on 2026-09-24, so every model: opus agent and CLAUDE_CODE_SUBAGENT_MODEL=opus already run on it) · sonnet = Sonnet 5 ($2/$10) · haiku = Haiku 4.5 ($1/$5). Live copy: claude-usage --modelos.
• Opus 5.5 notes: thinking is always on and effort is the only dial (medium default; high for hard bugs; xhigh/max only with a measured gain — it thinks more per turn than Opus 5 at the same level). Strongest gains: multistep coding in a real repo, code review, long analytical deliverables, reading charts/screenshots. For frontend work name the specific default patterns to avoid instead of "no generic look".
• Planning, decomposition, arbitration → director (Fable). Never use Fable in execution loops.
• Architecture, strategy, aesthetic judgment → opus heads. They produce briefs and verdicts, not implementations.
• Implementation volume → sonnet executors by roster default; for code work that must land (multi-file, tests must pass, prod repos) the plan may route the executor to Opus 5.5 — at ~half the tokens per task it costs about the same as Sonnet 5 and reviews better. Mechanical chores → ops-utility.
• Procedures and knowledge live in skills (.claude/skills/); spawn agents only for context isolation or parallelism.
• Pipelines are invoked as skills: /pipeline-sistema, /pipeline-landing-page, /pipeline-campaign, /pipeline-content, /pipeline-copy, /pipeline-brand, /pipeline-luxo, /pipeline-refino-ui.
• /pipeline-refino-ui (2026-09-25): refines an existing product's interface against the Apple HIG web digest (brain/design-library/principles/apple-hig-web.md, corpus via scripts/hig-fetch.mjs); tooling scripts/screenshot.mjs (routes, auth, labels), scripts/a11y-scan.mjs (axe + keyboard + touch targets) and scripts/shots-compare.mjs (before/after sheet). Stage 5 runs dev-builder on Opus 5.5. To run pipelines from a client repo, symlink .claude/skills/<name> and .claude/agents/*.md into ~/.claude/ (done on the primary workstation on 2026-09-25); the skill anchors every path to $HOME/Agent System.

Quality Gates

• Copy — mkt-editor: PASS / BLOCK / EXIT with named rollback agent. Max 2 revision loops, then escalate to the user.
• Functional — dev-qa: can BLOCK dev-ops deploys.
• Aesthetic — dsg-critic: APPROVE / REVISE / REJECT, scored against brain/design-library/principles/ with breakpoint screenshots. Under /pipeline-refino-ui the HIG consolidated checklist is scored as an extra lens and the verdict compares "depois" against "antes".
• Irreversible external actions (production deploy, campaign activation, bulk sends) require explicit human approval. Ad campaigns are created PAUSED, always (.claude/rules/ads-safety.md).

brain/ Ownership

Conventions

• System prompts, rules, and docs: English. Content deliverables: PT-BR by default (overridable per project pack).
• Handoffs between agents are files in brain/ops/handoffs/ — explicit state, not implied memory.
• Reuse before reinvent: check brain/design-library/patterns/ and brain/knowledge/ before creating anything new (enforced by the IDS hook).
• Secrets live in .env at repo root (kept in the folder, gitignored). Never inline keys in code, agents, or skills — reference process.env.* / $VAR.
• Approved deliverables in project packs are append-only: add [ADDED YYYY-MM-DD] sections, never overwrite.
• Every knowledge file carries a Last updated: header.

Deploys

Deploys that matter use a dedicated Vercel account via vercel --token "$VERCEL_TOKEN" (loaded from .env). Production deploys pass through dev-qa first.

Brand & Design Language (2026-08-19)

The Agent System is now its own brand — the central identity replacing the retired <BRAND> system-level brand (client brands unaffected). Direction: black & white, Apple-grade, liquid glass, refined UI/UX. Spec: brain/design-library/principles/agent-system-design.md. Accent A (system blue) and wordmark W1 were adopted — see the spec. Everything the system itself ships (status pages, stack map, internal tools) uses this language.

Stack Map (Obsidian constellation)

brain/MAP.md is the hub; brain/departments/*.md are the six branch MOCs (rule per branch: memory → manager → operators; clients are the outer ring); brain/agents/*.md are graph-visible cards mirroring .claude/agents/ (hidden folders are invisible to Obsidian). Open the repo folder as a vault → graph view renders the constellation. Keep the cards in sync when the roster changes.

Client packs

Each client or company lives in its own pack under brain/projects/<name>/, created from brain/projects/_template/. Packs are private and are not part of this repository. Agent memories live under .claude/agent-memory/<agent>/ (private too — the folders ship empty).