An always-on autonomous strategy council of AI co-founders. You give it a hard question; a panel of seven deliberately-different expert personas debates it on a schedule, votes, extracts the follow-up work, executes that work through a task system with real dependency management and peer review, and records everything in a linked knowledge vault that feeds the next day's deliberation. A human is in the loop only where it genuinely matters.
This is the open-source engine behind a private system I built and run (codename Founders Circle). It ships the orchestration, the seven persona templates, and an empty vault scaffold that populates as you use it. All real company data — deliberations, decisions, financials, prospects, credentials — has been removed. It is a working reference and a portfolio piece, not a turnkey product. — Sage Dutra
- A council tuned to disagree. Seven personas, each with a real business-strategy framework lens and a sharp bias — a skeptical market analyst, a ship-it-in-three-weeks builder, a distribution-obsessed growth lead, a cold-eyed finance officer, a category-creating creative, a chief-of-staff synthesizer, and Patron, who argues in the first person as the customer. The friction is the point. Maestro synthesizes and tallies a domain-weighted vote; it never grades its own work.
- Two orchestration layers, honestly documented. A durable Temporal foundation (workflows with typed timeouts, heartbeats, signal/query handlers, classified failures) and the Parthenon operating layer that actually runs day to day — bash + schedulers spawning fully-tooled headless Claude Code sessions as agents, with deliberate per-round model switching for cost. The two-layer story, and why it evolved that way, is written up plainly in
docs/FC-SYSTEM-OVERVIEW.md. - A real task system, not a to-do list. Postgres-backed, with an 8-state lifecycle, optimistic claim-and-release locks (two runners can never double-execute), a dependency cascade that auto-unblocks downstream work, template-variable injection of dependency outputs, a peer review/revision loop, and a two-phase planning gate for complex tasks.
- Structural anti-fabrication. Source credibility tiers and a confidence gate on research; a data-integrity gate that tags
[UNVERIFIED]numbers; a requirements check that blocks any deliverable whose facts aren't sourced; a nightly audit that checks active work for drift. The system is built, top to bottom, to refuse to invent data. - A compounding knowledge vault. Every round, idea, decision, and learning becomes an individual, tagged, wikilinked note with a typed frontmatter schema. Institutional memory compounds instead of evaporating at the end of each chat.
Topic enters the hopper (you, or a queue)
|
Parthenon: 7 agent rounds/day refine it, on a schedule
|
Closing round: Maestro votes, synthesizes, EXTRACTS TASKS
|
Workshop: claims tasks, injects dependencies, runs the assigned agent
|
Peer review vs. acceptance criteria -> approved -> lands in the vault
|
Dependency cascade auto-unblocks downstream tasks
|
Nightly audit: drift check, stale check, workload rebalance
|
Vault notes feed the next day's deliberation -> (loop)
Three "rooms" coexist: Parthenon (formal deliberation), Workshop (execution), and a deliberately task-free Greenroom (casual). The interface is Discord; there is no custom dashboard.
| Agent | Role | Framework lens |
|---|---|---|
| Maestro | Chief of Staff | Cross-framework synthesis; tallies the weighted vote, routes escalations |
| Pulse | Market Analyst | Porter Five Forces + Christensen Disruption; credibility tiers, confidence gate |
| Rivet | Builder | Drucker Effectiveness; "is this shippable in 3 weeks?" |
| Venture | Growth Lead | Godin Purple Cow + Kim/Mauborgne Blue Ocean; flywheels over paid acquisition |
| Abacus | Finance Officer | Collins Flywheel + Taleb Antifragile; unit economics, data-integrity gate |
| Rembrandt | Creative Director | Meadows Systems Thinking + Doumont; category creation |
| Patron | Prospect Voice | First-person customer simulation; "would I actually respond to this?" |
Each persona is a template directory (src/agent-templates/<name>/) with an identity, a personality, a framework, hard rules, output standards, and skills. Editing a template changes every future run. The personas here are illustrative; swap them for your own company's council.
- TypeScript (strict) + Node 20+
- Temporal.io for the durable workflow foundation
- PostgreSQL for council state + the task system
- The Claude Code CLI (
claude -p) as the reasoning runtime — each agent is a disposable, fully-tooled headless session booted with a persona - Google Gemini Flash for cheap high-volume generation, Tavily for research
- Discord (
discord.js) as the entire human interface - An Obsidian vault (Markdown + YAML frontmatter) as the knowledge layer
The engine, the persona templates, the task-system schema, the Parthenon and Workshop runners, the audit, and the vault schemas are all real. What ships here is scaffolded for a fresh deployment: the vault is empty (it fills as you run), the configs are .example files you copy and fill, and running it for real requires your own Postgres, a Claude Code CLI on PATH, Gemini/Tavily keys, and a Discord bot. See docs/SETUP.md. Nothing here calls out to a paid service until you provide credentials.
| Doc | What's in it |
|---|---|
docs/FC-SYSTEM-OVERVIEW.md |
The full system explainer — architecture, the two layers, coordination, honest caveats |
docs/SETUP.md |
Prerequisites and how to run it |
docs/plans/BUILD-PLAN.md |
The original spec-first build plan |
docs/plans/AGENT-OFFICES-PLAN.md |
The task-system design |
docs/plans/OBSIDIAN-VAULT-PLAN.md |
The knowledge-vault design |
vault/README.md |
The vault folder taxonomy + frontmatter schemas |
Apache License 2.0. Copyright 2026 Sage Dutra. See NOTICE.