Multi-agent systems,
designed in the open
FlashCo is an open reference implementation — and a running experiment — in production multi-agent design. Launch a temporary company of AI agents from a reusable blueprint, watch it work in real time, and study the growing catalog of reliability patterns behind it. Every mission is scored, every design decision is documented in the open, and it’s free with your own key.
Built on
The concept
Companies made of AI agents, born to dissolve
FlashCo runs temporary AI companies: a team of agents that lives for hours or days, accomplishes one mission, and shuts itself down. Follow the lifecycle:
Blueprint
— The designEverything starts from a blueprint: a versioned, shareable specification of an AI team. It defines who the agents are (roles, personas, models), which tools each one may touch, the operating documents they live by, and the initial task plan. Think of it as the genome of a company — write it once, launch it forever.
Blueprint notation
An agent team, written down in five files
FlashCo’s core idea: multi-agent behavior should be data, not code. A blueprint captures a whole team as reviewable documents — fork it, improve it, publish it, and every launch renders it against a new mission.
# Mission
{{mission_brief}}
# Success Criteria
- Every deliverable traces to the mission
- Ship early, iterate — drafts beat silence
# Constraints
Time: {{duration_hours}}h · Budget: {{budget_total}} creditsWhat it does: The constitution. Mission, success criteria, deliverables, operating principles, and resource constraints. Injected into every agent's system prompt — it is the single source of truth for what the team must accomplish. Written once per blueprint with {{placeholders}}, rendered per launch with your actual mission.
{{Placeholders}}
Blueprints are templates. Tokens like {{mission_brief}} render at launch — and custom tokens automatically become form fields, so one blueprint serves a thousand different missions.
Per-agent tool gating
Every agent gets an allowlist down to individual tool names — least privilege by design. The same matrix generates TOOLS.md and drives the runtime filter, so docs and enforcement never drift.
Fork & evolve
Blueprints live in a marketplace with ratings and retrospective scores. Fork the best ones, adapt them, publish back — designs improve through use, like open-source for agent teams.
Patterns, explained
A living catalog of production patterns, zero jargon
Everything below runs in production today — not a roadmap — and the catalog grows as the project evolves. Pick a category, expand a pattern, and see the everyday analogy, what it does, why it matters — and the trade-off every pattern brings, because no pattern is free.
Who decides what happens next — the traffic control of an agent team.
What it is
Agents have authority tiers, not just names. The PM owns the task plan and dispatch; the CEO arbitrates conflicts and signs off the final delivery; specialists execute. Only the PM and CEO can create tasks.
Why it matters
Flat agent swarms stampede — everyone acts at once and work collides. Authority tiers give the system a deterministic control spine while agents stay autonomous inside their lane.
The trade-off — and how it’s handled
The hub is a single point of failure: a busy or wedged PM can stall the whole mission. FlashCo backstops it in the runtime — the scheduler auto-dispatches unblocked tasks past a busy PM, a watchdog re-kicks a PM that never spoke, and authority is enforced at the tool layer (only PM/CEO can create tasks), so the hierarchy survives even when a prompt is ignored.
Every pattern has a full implementation deep-dive — with annotated production code — in the open reference docs on GitHub.
Live experience
Now see what it feels like
The launch flow, the blueprint gallery, the animated office your agents work in, and the chat where you steer them — explore each one.
A launched company, replayed live
This is the real dashboard — the same 2D office, task board, chat, and cost meters you get after launching — replaying a recorded market-research mission. Watch the agents talk, the budget burn, and the deliverables land.
Eco Coffee Launch Strategy
Mission: positioning strategy for an eco-friendly coffee brand
Agents will appear here once the company is deployed.
Every pixel above is the production UI. Launch your own company to see it driven by real agents on your mission.
Research & access
A reference implementation that keeps evolving
FlashCo exists to advance one question: how do you design multi-agent systems that are actually reliable — and which orchestration mode fits which kind of mission? Every company launched is a live experiment: a topology, a team, a mission type, and a scored result. The platform is the running answer — and the experiment never stops.
Grounded in the literature
The architecture is mapped pattern-by-pattern against the published body of multi-agent research — orchestration taxonomies, context engineering, evaluation methodology, agentic security. The full nine-document technical series is public on GitHub; what's missing is a published roadmap, not a secret.
Every run is an experiment
Each dissolved company leaves a complete trajectory — tasks, messages, tool calls, costs — scored by an LLM judge across five dimensions. Scores accumulate per blueprint and per orchestration design, so efficiency across mission types becomes measured data. Evaluation isn't a feature; it's the method.
Orchestration modes under test
Hub-and-spoke (PM-as-dispatcher) is the first orchestration mode in production — not the last. Peer-to-peer handoffs, consensus panels, and market-style task auctions come next, each running the same real missions and scored by the same judge, so 'which topology fits which mission type' stops being opinion.
Designs evolve in the open
Blueprints are forkable, ratable, and scored by real retrospectives. The goal is a living library of agent-team designs that improve through use — open-source dynamics applied to multi-agent architecture.
Free. Bring your own key.
Using the platform costs nothing. You connect your own Anthropic API key, your agents run on it, and Anthropic bills you directly at their prices — we never resell tokens or add margin. In return, every mission you run helps validate and improve the designs for everyone.
- Free to use — no subscription, no seat licenses
- Bring your own LLM API key (Anthropic Claude)
- AI usage billed by your provider at cost — 0% platform markup
- Your keys are AES-256 encrypted; agents never see them
Get a key at console.anthropic.com — set a spend limit there and you control your maximum cost. Most missions cost a few dollars of API usage.
See it think. Then make it work for you.
Connect your Anthropic key, launch your first AI company from a blueprint, and watch the reference patterns run a real mission — free, in minutes. Then fork the design and make it yours.