Founders Circle.A council that runs itself.
Seven AI agents with distinct strategic framework lenses, scheduled to meet every day, deliberate one topic across seven rounds and fourteen hours, vote, extract tasks, and ship those tasks back to themselves. The only interface is Discord. The institutional memory is a tagged Obsidian vault.
Autonomous multi-agent deliberation in production. Not a prompt with personas. A coordination layer with voting, dependency tracking, source-credibility gates, and a hard refusal to ship unsourced material.
What Founders Circle Is.
Founders Circle is an autonomous strategy council. Seven AI agents, each with a distinct framework lens, meet on a schedule and produce strategic output: positioning memos, pricing models, prospect selection criteria, messaging specs. The council debates one topic per day, persisting up to seven days, with seven rounds of refinement spaced every two hours. At the end of each topic, the synthesizer agent votes and extracts tasks. Other agents claim and ship those tasks.
Three things define the system. It is advisory, not operational. The agents write documents. They never write code, never touch the FORGE pipeline, never contact clients. It runs without prompting. The council convenes on a clock; I do not have to start a session. And it is a closed loop. Discussions produce decisions, decisions extract tasks, tasks get shipped and reviewed, finished work lands in a knowledge vault, and the vault feeds the next discussion. Institutional memory compounds instead of evaporating at the end of each chat.
The mental model: a seven-person executive team that costs API tokens instead of salary, works fourteen hours a day, never forgets a prior decision, and escalates to me only when it genuinely needs a human call.
Seven Agents. Seven Lenses.
The friction between them is the point. They are tuned to disagree productively.
Maestro
Cross-framework synthesis.
The synthesizer, not a voter. Speaks last, fewest words. Breaks 3-3 ties only after exhausting clarification. Tallies votes with domain-weighted scoring. Routes escalations to me.
Pulse
Porter Five Forces + Christensen Disruption.
Deeply skeptical. Demands search volume. Source credibility tiers T1 through T4. Confidence scoring 0.0 to 1.0 with a hard gate at 0.6. Treats unvalidated assumptions as lies.
Rivet
Drucker Effectiveness.
Brutally pragmatic. Asks one question: is this shippable in three weeks? Kills beautiful ideas that cannot deploy. Prefers event-driven over polling, simple over clever.
Venture
Godin Purple Cow + Kim/Mauborgne Blue Ocean.
Distribution-obsessed. Hates paid acquisition. Thinks in flywheels and referral loops. Guards the core growth motion.
Abacus
Collins Flywheel + Taleb Antifragile.
Coldly rational. Tracks unit economics. Data integrity gate: tags unverified numbers as unverified. Will not approve plans without revenue math.
Rembrandt
Meadows Systems Thinking + Doumont Communication.
Provocateur. Hates safe positioning. Pushes category creation. Always runs on the Opus model for maximum creative bandwidth.
Patron
First-person prospect simulation.
Speaks as the actual client. Emotionally honest, has been burned by agencies before. Tests every message against the question "would I actually respond to this?" Never breaks character.
Decision weighting (Maestro's tally): Demand 20% · Revenue 20% · Competition 15% · Technical 15% · Agent Fit 15% · Prospect Reality 15%. Weights shift by decision type. A pricing decision elevates Abacus. A technical decision elevates Rivet.
Two Layers. By Design.
The system was built on Temporal as the durable workflow foundation. The daily operating layer that emerged is a set of bash scripts and macOS LaunchAgents. Both layers exist. Each handles what it's best at.
Temporal.io is the durability foundation. Workflows, retries, crash recovery, signals, event history. The task queue is isolated from FORGE's task queue so the two systems cannot interfere with each other. Temporal is what makes "the council convenes every day at 8 AM" a guarantee, not a hope.
The Parthenon operating layer is what actually runs the daily cycle. A bash script (parthenon-round.sh) is scheduled by LaunchAgent every two hours. It selects the topic for the day, picks the agent whose lens fits the current round, spawns a headless Claude Code session with that agent's personality prompt, captures the output, and posts it to the Discord channel. At 10 PM, Maestro runs the closing round: tallies votes, extracts tasks, opens new Obsidian notes.
The Workshop layer is the task system. Tasks extracted from Parthenon discussions land in a Postgres table. Other agents claim them via an optimistic locking pattern. They produce a draft. A peer agent reviews. The original drafter revises. Once an artifact passes review, it goes into the curated knowledge vault. Finished work compounds.
A separate Maestro audit cycle runs at 9 PM daily. It detects stale tasks, looks for drift between active decisions and live domain documents, and rebalances workload. The audit is what keeps the system from accumulating debt.
The Pieces.
The Judgment Calls.
A multi-agent council is just a group chat unless the coordination patterns are deliberate. These are the patterns that keep this one honest.
Source credibility, not vibes.
Pulse, the market analyst, scores every claim against a four-tier credibility rubric (T1 primary sources, T2 reputable secondary, T3 industry commentary, T4 marketing material). Anything below a 0.6 confidence gate gets flagged as unverified and refuses to ship. The council structurally cannot produce a memo built on assumptions.
Data integrity gate on every number.
Abacus, the finance officer, applies the same gate to financial claims. Any number without a verifiable source is flagged. No revenue projection ships without unit economics math attached. Vibes-based forecasting is structurally impossible.
Patron speaks as the actual client.
The seventh agent, Patron, is a first-person prospect simulation. Whenever the council debates messaging, pricing, or outreach, Patron tests the proposal against "would I actually respond to this?" The agent never breaks character. The result is that customer reality is in the room, not just inferred from data.
Optimistic claim-and-release locking on tasks.
When an agent claims a task, the lock is optimistic. The agent can hold it, work it, or release it back to the pool. Dependencies cascade automatically: if a task unblocks another, the downstream task becomes claimable. A two-phase planning gate runs ahead of complex work: a peer agent reviews the plan before execution begins.
Per-round model switching as cost discipline.
Deep reasoning rounds use Opus. Refinement rounds use Sonnet. Brainstorm fan-out uses Gemini Flash. This is not a free-for-all; the model for each round is pinned in the workflow definition. A typical full day's deliberation costs less than a single Opus call on a long prompt.
The 9 PM audit.
Every night at 9 PM, Maestro runs an audit cycle. Stale tasks get surfaced. Drift between active decisions and live domain documents is detected. Workload is rebalanced across agents. The audit is what prevents the system from accumulating decision debt the way a busy team accumulates Slack threads.
The Pattern Travels.
The Founders Circle pattern (scheduled deliberation, framework-driven agents, voting with domain weighting, source-credibility gates, institutional memory in a tagged vault) generalizes well beyond a solo agency.
The most direct lift is a fractional executive team for small businesses that cannot afford a real one. Vertical persona packs (dental, legal, trades, e-commerce, restaurants) swap the framework lenses without changing the coordination layer. A second lift is institutional memory infrastructure: the same Obsidian vault pattern, with the same proposal-then-curate gate, works as a decision-of-record system for any team where context evaporates between Slack threads. A third lift is white-label fractional CFO and CMO firms, who could embed an always-on AI council inside their advisory engagement as a retention layer.
The architectural piece worth flagging separately: the two-layer split, Temporal as the durability foundation and bash scripts as the operating layer, is honest about how the system evolved and is worth reading as evidence of pragmatic engineering judgment. The bash layer is not the elegant choice. It is the choice that works on a Mac Mini today and migrates to Temporal-native workflows when the cost-benefit changes.
There's More.
Founders Circle is one of the AI systems I've built. The resume supplement page has the full set, including FORGE (the centerpiece) and the Solution Factory.