Agent Management Architecture — Soilytix + Personal¶
Reference document for AI agent department gestão em Soilytix + Cleiton's personal projects. Filed Apr 28 2026 antes do Reporting Agent V3 build (May Wk 1-2). Re-read antes de Phase 2 trigger (May Wk 3+) e antes Phase 3 review (Set-Out 2026).
TL;DR¶
Agent management não é uma coisa só — são 8 layers com tools diferentes. Para Soilytix (multi-stakeholder, EU GDPR, low-budget) o stack ideal Apr 2026 é managed SaaS-first (não self-host até Cleiton evoluir DevOps comfort). Para personal track (xquads-evolved), os mesmos primitives reusam mas com governance light. 3 sistemas docs/tasks ficam como are: Notion (strategic) + GitHub (technical) + Obsidian wiki (knowledge synthesis). Linear, Cursor, Paperclip, Multica off the table até signals concretos. Plano: ship Reporting Agent V3 May Wk 1-2 com observability wired desde Day 1, Phase 2 trigger May Wk 3+, 8-agent department target Out 2026.
🧬 Os 8 layers de agent management¶
Antes de escolher tools, decompor o problema:
| # | Layer | O que decide | Soilytix critical | Personal critical |
|---|---|---|---|---|
| 0 | Identity | Quem é cada agent? Que credenciais? Audit trail? | ⭐⭐⭐ (multi-stakeholder + GDPR) | ⭐ (só Cleiton) |
| 1 | Runtime | Onde corre? Cron? Interactive? VM? | ⭐⭐ (cron + reliability) | ⭐⭐ (interactive + experimental) |
| 2 | Memory | Lembra entre runs? Cross-session? Cross-agent? | ⭐⭐ (week-over-week reports) | ⭐⭐⭐ (continuidade pessoal) |
| 3 | Communication | Como fala com humanos? Outros agents? APIs? | ⭐⭐⭐ (Notion/Slack/HubSpot) | ⭐ (Slack/files) |
| 4 | Observability | O que fez, porquê, quanto custou, falhou? | ⭐⭐⭐ (cost discipline + audit Bruno) | ⭐⭐ (curiosidade + cost) |
| 5 | Coordination | Task queue. Assignment. Dependências. | ⭐⭐ (Phase 2+) | ⭐ (single-task usually) |
| 6 | Governance | Budgets. Approval gates. Kill switches. | ⭐⭐⭐ (Bruno visibility + €) | ⭐ (só carteira tua) |
| 7 | Skills/Capability | Reusable patterns. Prompt library. Tools. | ⭐⭐ (3 candidate skills filed) | ⭐⭐⭐ (xquads-squads existe) |
Insight crítico: os tools NÃO competem entre si — endereçam layers diferentes:
- Langfuse = Layer 4 (observability)
- Paperclip = Layer 6 (governance) + parcial 5 (coordination)
- Multica = Layer 5 (coordination) + parcial 7 (skills)
- gstack (Garry Tan) = Layer 7 (skills) + parcial 1 (runtime patterns)
- Notion = Layer 3 (communication) + parcial 5 (coordination)
- gbrain = Layer 2 (memory)
- LiteLLM = Layer 1 (runtime — model routing)
- Linear = Layer 5 (coordination) + parcial 3 (comms)
- Cursor = development tool, NÃO endereça nenhum layer agent management
Vais usar vários, não um.
🌱 Soilytix track — multi-stakeholder, GDPR, revenue-critical¶
Constraints únicos¶
- 3 stakeholders: Cleiton (Head of Revenue), Bruno (CEO, results-driven, não dashboard-deep), Julien (CCO)
- GDPR + CSRD (EU) — data residency hard requirement
- Revenue-critical — failures = perda de €
- Budget low-radar — < €30/mo Phase 1+2 sob radar Bruno
- Cleiton DevOps comfort: SaaS-only por agora, evolve para self-host Q3+
Stack final (managed-only, Apr 2026 decision)¶
| Layer | Tool | Cost/mês |
|---|---|---|
| 0 Identity | Shared OAuth (Cleiton) + per-agent Slack apps Phase 2+ | €0 |
| 1 Runtime | GHA cron (free) + LiteLLM as Python lib (no separate VM) | €0 |
| 2 Memory | Notion DB row "agents" + Cockpit Sheet | €0 |
| 3 Communication | Notion (Commercial Board) + Slack #soilytix-agent | €0 (existing) |
| 4 Observability | Langfuse Cloud EU free tier (50k obs/mo) | €0 |
| 5 Coordination Phase 1 | Notion DB | €0 |
| 5 Coordination Phase 2+ | Notion DB extended OR Linear free (250 issues) | €0 → €8/mo Phase 3 |
| 6 Governance | Anthropic budget alert + monthly review email Bruno | €0 |
| 7 Skills | GitHub agents/skills/ markdown + xquads converted |
€0 |
| API spend Phase 1 | Anthropic + Gemini Flash | €2-15/mo |
| API spend Phase 2 | + 2-3 agents | €30/mo |
| API spend Phase 3 | 5+ agents | €60-80/mo |
| API spend Phase 4 | 8 agents department | €100-150/mo (precisa Bruno report) |
| Hetzner VM (post Q3) | Quando Cleiton evolve DevOps comfort | €4.5/mo |
Output design — Bruno-friendly (Q1: results-driven)¶
Não é dashboard. É Slack DM weekly Friday + immediate alerts:
📊 W19 Traction Friday Report
ROAS: 4.2× (+0.5 vs W18) ✅
Pipeline: +€7,500 weighted (3 new deals)
Demos booked: 18 (+4)
Leads: 42 (-2, monitoring)
Top alert: nenhum esta semana
Full report: <Notion link>
Quando algo anómalo (W3 spike-style):
🚨 LinkedIn CPL spike — €165 (+106% vs benchmark)
Action plan filed: <link>
Hipóteses: fatigue OR attribution lag
Per-phase timeline (compressed Apr 2026 decision)¶
Phase 1 — May Wk 1-2 — Ship V3 + start narrative¶
Build: - Reporting Agent V3 — single agent, Stack C managed - Anthropic Sonnet 4.6 (narrative) + Gemini 2.5 Flash (reads) + Haiku 4.5 (anomaly) - LiteLLM as Python lib (não VM separate) - GHA cron orchestration + Langfuse Cloud EU traces wired Day 1 - Notion + Slack outputs
Cost real Phase 1: ~€2-5/mo
Narrative (build-in-public):
- LinkedIn post sobre o build
- Soilytix blog post (uses Blog Pipeline GHA shipped Apr 27)
- Adicionar a soilytix/content/blog-drafts/
Phase 2 — May Wk 3 → Jun — Compress Jun-Aug original plan¶
Build: - Agent #2: Reply Triage (Haiku 4.5 classify replies) - Agent #3: Lead Enrichment ou Content Reviewer - Custom Streamlit dashboard 1 day — agents alive + cost per agent + last-run + kill-button (replaces Multica/Paperclip self-host need)
Decision points: - Notion API rate limit começou a partir? → migrar tickets para Linear free - Senão → mantém Notion DB
Narrative: LinkedIn weekly + 1 podcast / livestream pitch
Phase 3 — Jul-Aug — Scale to 5 agents + framework emerges¶
Build: - Agents 4-5 - Personal track parallel: xquads → SKILL.md migration - LiteLLM proxy passa Hetzner VM (€4.5/mo) quando Cleiton evolve B-level
Narrative + Framework: - "Soilytix Stack Open Sourced" — abstract patterns em framework reusable, MIT license, separate repo - Speaking gigs (DLG, Agritech meetups, AI conferences EU)
Phase 4 — Set-Out — 8 agents + cement authority¶
Build: - 6-8 agents production - Personal projects on same patterns (proves replicability) - Framework GitHub repo: 1k+ stars target
Narrative + Career: - 1-2 keynote talks / podcasts - Wiki público (subset cleitonOS-workspace) - Soilytix promo discussion (cabe a Bruno)
🧪 Personal track — solo, exploratório, memory-rich¶
Constraints únicos vs Soilytix¶
- 1 stakeholder (Cleiton)
- Sem GDPR pressure
- Failures = oportunidade aprender
- Budget pessoal flexível
- Existing investment: xquads-squads (13 squads markdown), ti-na-europa, autoresearch-mlx, Obsidian, MCP custom servers
Asset escondido — xquads-squads¶
Tens 13 squads markdown em xquads-squads/. Hoje são static documents. A migração óbvia: transformar em executable SKILL.md format estilo gstack:
---
name: hormozi-offer-stress-test
description: Apply Hormozi value equation framework to stress-test an offer
allowed-tools: [Read, Write, AskUserQuestion]
triggers:
- "stress test this offer"
- "is this offer good"
---
## Forcing questions
1. Dream Outcome: que resultado específico cliente atinge?
2. Perceived Likelihood: que evidência tens que isto vai funcionar?
3. Time Delay: quanto tempo entre purchase e value?
4. Effort & Sacrifice: o quanto cliente tem de fazer?
## Anti-patterns
- Vague outcomes ("save time")
- Hypothetical testimonials
- Claims without proof
## Output
Score 0-10 + 3 specific revision suggestions.
Resultado: tens personal gstack equivalent com tua wisdom curada.
Stack por layer — Personal¶
| Layer | Tool | Por quê |
|---|---|---|
| 0 Identity | Cleiton OAuth + 1 GitHub PAT + 1 Slack PAT | Single-user, não vale per-agent |
| 1 Runtime | Claude Code interactive (default) + GHA cron | Interactive para experimentação, GHA recorring |
| 2 Memory | gbrain (Garry Tan) cross-machine + Obsidian wiki | Cleiton M4 + Mac empresa partilham contexto |
| 3 Communication | Telegram bot @StealthOps + Obsidian | Mobile + deep work |
| 4 Observability | Langfuse Cloud EU free (compartilhada com Soilytix se quiseres) | Free, EU, dual-purpose |
| 5 Coordination | xquads-squads/ + GitHub issues per project | Não vale Multica para 1 user |
| 6 Governance | Anthropic console budget alert mensal | Disciplina mental, não tool |
| 7 Skills | xquads → SKILL.md format + clone gstack patterns | 13 squads → executable |
Action concreta Personal (próximas 4 semanas, post-Soilytix Phase 1)¶
Não competes com May Wk 1-2 (Soilytix prioridade). Personal track arranca May Wk 3+:
Week 1 personal (~30-60 min weekend)¶
- Clone
garrytan/gstacklocalmente - Estudar 3 SKILL.md (office-hours, plan-ceo-review, design-shotgun)
- Pick 1 xquad — recomendo hormozi-squad (aplicável a Soilytix copy também)
- Converter em
xquads/skills/hormozi-offer-stress-test/SKILL.md - Test: invocar via Claude Code, verificar framework é seguido
Week 2-3 personal¶
- Converter 4-5 squads top-priority
- Recommended order:
- hormozi-squad (offer/copy review)
- copy-squad (LinkedIn posts, outreach)
- claude-code-mastery (meta — agents melhores)
- advisory-board (decision-making)
- storytelling (content)
Week 4 personal¶
- Setup gbrain self-host (Hetzner VM mesma instance se Cleiton evolved B-level, ou separate)
- Cross-machine memory: M4 home + Mac empresa share contexto
Mês 2-3 (Jun-Jul)¶
- Ti-na-europa: aplicar mesmo pattern —
agents/folder skill library - Telegram bot: wire skills via Claude Code subprocess (já tens shadow-mode)
🤝 Primitives partilhados — build once, use both¶
| Componente | Soilytix usa para | Personal usa para |
|---|---|---|
| Hetzner CX22 VM €4.5/mês (post Q3) | LiteLLM dedicated + Langfuse self-host | gbrain server + same VM |
| Langfuse Cloud EU (Phase 1-2) | Reporting Agent traces + cost | xquads runs traces |
| gbrain memory (Q3+) | Cross-machine Cleiton state | Memory cross-session entre projetos |
| GitHub Actions free tier | BD/PR/Blog/Cockpit pipelines | Personal cron (research, content gen) |
| Anthropic API key Cleiton pessoal | Phase 1 V3 (€5-7/mo) | Personal experiments |
| Skill library pattern markdown SKILL.md | Soilytix agents/skills/icp-stress-test/SKILL.md |
xquads transformados |
Big insight: mesma VM Hetzner serve ambos quando Cleiton evolui B-level (Q3+). €4.5/mês cobre LiteLLM + Langfuse self-host + gbrain. Personal e Soilytix compartilham infrastructure, mantêm identity separate (databases isolados, secrets diferentes).
📚 Documentation stack — 3 systems naturais¶
Linear off the table. Cursor não cabe. Stack actual stays:
Strategic / Business / Tickets
└─→ Notion (Commercial Board)
tasks, decisions, ICP analysis, deal pipelines, OKRs
Technical / Code / Stack
└─→ GitHub (repo READMEs, wikis)
soilytix-revenue-automation/README.md
per-repo docs, commit-history-driven
Knowledge / Synthesis / People / Concepts
└─→ Obsidian wiki (cleitonOS-workspace/wiki/)
frameworks, gurus, concepts, master-map
Karpathy-style 2nd brain
Cada um tem natural fit: - Bruno (CEO) lê Notion — vê pipeline, decisões, briefs - Engineers lêem GitHub READMEs — vê deploy, contribuir, debug - Cleiton sintetiza em Obsidian wiki — synthesise across all sources
Triggers para revisitar Linear (Out 2026+): - Notion API começar a partir agents (rate limit hits weekly) - Equipa cresce 5+ people active em tickets daily - Velocity-driven workflows (sprints rigorous) em vez de strategic-driven
❌ Tools considered + rejected (Apr 2026)¶
Cursor (cursor.com)¶
- Categoria: AI-powered IDE + cloud agents para coding
- Por que rejeitado: wrong layer. Não endereça layers 0-7 agent management. Switching cost real (Cleiton já tem Claude Code CLI muscle memory). No EU residency stated → GDPR risk se touches customer data.
- Quando revisitar: Bruno (CEO) ou future eng hire para product code, não Cleiton GTM ops.
Paperclip (paperclip.ing)¶
- Categoria: Layer 6 governance + parcial 5 coordination, "AI agents as employees" metaphor
- Por que rejeitado AGORA: self-host first → Cleiton C-level DevOps comfort, não vai correr VM Phase 1+2.
- Quando revisitar: Q3+ quando Cleiton evolve B-level OR Paperclip lançar Cloud SaaS
Multica (multica.ai)¶
- Categoria: Layer 5 coordination + parcial 7 skills
- Por que rejeitado AGORA: mesma razão que Paperclip — self-host required.
- Quando revisitar: Q3+ ou Cloud lançar
Linear (linear.app)¶
- Categoria: Layer 5 coordination (tickets) + parcial 3 comms
- Por que rejeitado AGORA: migration cost Notion → Linear = sprint inteiro. Bruno + Julien já lêem Notion. Notion API rate limit ainda não nos parteu.
- Quando revisitar: Out 2026+ se Notion API rate limit hits weekly OR equipa cresce 5+ users
Bedrock AgentCore (AWS)¶
- Por que rejeitado: lock-in AWS desnecessário. Não estamos em AWS.
Microsoft Copilot Studio¶
- Por que rejeitado: $200/mo min tenant. Caro. Lock-in Microsoft.
OpenAI Agent Builder / Swarm¶
- Por que rejeitado: lock-in OpenAI. Stack já é Anthropic + Google + Meta (LiteLLM router).
🎯 First 3 actions to start IMMEDIATELY (May Wk 1)¶
1. Setup Langfuse Cloud account EU region (15 min)¶
- https://cloud.langfuse.com — escolher EU region durante signup
- Get API keys + project ID
- Add
LANGFUSE_PUBLIC_KEY+LANGFUSE_SECRET_KEYGHA secrets - Wire em Day 1 do Reporting Agent V3 build (não retrofit depois)
2. Setup Google AI Studio free tier Gemini API (10 min)¶
- https://aistudio.google.com — free tier generoso
- Get API key
- Add
GOOGLE_AI_API_KEYGHA secret - LiteLLM Python lib config: route reads através disto
3. Decide narrative cadence (10 min mental + commitment)¶
- LinkedIn post: weekly Friday build-in-public, 1 specific learning per week
- Soilytix blog: monthly comprehensive case study (uses Blog Pipeline GHA)
- Wiki public excerpt: quando 3-month milestone (Jul)
Total setup time: 35 min real work. Tudo manual, sem Hetzner, sem Docker.
⚠️ Reality check honesto — Q5 success em 6 meses¶
Cleiton's stretch goal Out 2026: 8 agents + tech authority + C-level promotion + internet authority + rich.
Probability per outcome:
| Outcome | Probability achievable em 6 meses | Comment |
|---|---|---|
| 8 agents running | 65% | Stretch mas possível com aggressive timeline + framework reuse. Mais provável 5-6 agents reais + 2-3 stub. |
| Tech authority on AI-native | 40% | Depende de content discipline (1 LinkedIn post/week minimum). Sem isso, zero. |
| C-level promotion Soilytix | 30% | Depende Bruno calibration + Soilytix revenue trajectory + funding round. Não controlas tudo. |
| Internet authority | 35% | 6 meses é curto. Precisa 1-2 viral pieces + sustained content. |
| Rich from personal | 15% | 6 meses é cedo demais para revenue real de personal projects. 18-24 meses mais realista. |
Honest framing: ambition is healthy AND all simultaneously em 6 meses é unrealistic. Pick 2-3 prioritários:
- 8 agents Soilytix (control)
- Tech authority (controlable via content discipline)
- C-level / wealth = downstream effects, deixa happen naturally
🔓 Migration safety — não lock-in¶
Princípios para escolhas hoje não viram prisão amanhã:
- Skills em markdown SKILL.md format — portável entre Claude Code, gstack, Multica, custom runtime
- Memory em Postgres standard — gbrain SQLite/Postgres simples, exportável
- Traces em OpenTelemetry format — Langfuse exporta, qualquer observability importa
- Notion dados exportáveis — Markdown export nativo
- Tudo self-host via Docker Compose (eventual) — VM swap não corrompe stack
Anti-pattern a evitar: adoptar SaaS que só exporta para si mesmo (Bedrock AgentCore proprietary state, OpenAI Agent Builder schemas opaque).
🤔 What's uncertain (open questions)¶
- Paperclip vs Multica long-term winner: ambos lançados Apr 2026 (novos). Pode ser que ambos morram e algo melhor emerja Q3.
- gbrain stability: 12k★ mas Garry Tan move-fast. Pin versão concreta + backups disciplined.
- Soilytix Phase 3 reality: o "8-agent department" é vision. Phase 2 (3-5 agents) pode descobrir que 4 chega. Não over-build.
- Migration cost xquads → SKILL.md: estimei 30 min/squad, pode ser 2h se queres polish nível-gstack. Comece com 1, calibra.
- Notion as Layer 3 tem teto: 3 req/s API rate limit vai partir aos 5+ agents. Pode forçar migration para Linear cedo.
Cross-links¶
- Wiki:
thin-harness-fat-skills— pattern arquitectural por trás - Wiki:
ai-native-company— Diana YC framework (closed loops, queryable org) - Wiki:
20x-companies— case studies validating approach - Wiki:
tools/g-stack— Garry Tan implementation - Wiki:
tech/REPORTING-AGENT-V3— concrete Phase 1 build spec - Wiki:
people/garry-tan/profile - Notion: SOI-CONSTITUTION-00 — SDD adoption ticket
- Notion: MKT-OPS-04 Reporting Agent V3 — build tracker
Sources¶
- Conversation Cleiton ↔ Claude Code, Apr 28 2026 (deep planning session)
- Subagent research Apr 28 2026 — landscape Apr 2026 + free tier verification Perplexity
- Direct WebFetch paperclip.ing + multica.ai + cursor.com/pricing Apr 28 2026
- knowledge/tech/2026-04-28-* (3 YouTube videos: Garry Tan, Diana YC, "20X companies")