Skip to main content

​
You’re the bottleneck, not your agents

Research Preview — We think orchestration is the missing piece between “AI can code” and “AI can ship.” This is experimental software. APIs will change, things will break, and we’re building in public because that’s the only way to build something real. Join us on Discord to argue about it.

​
The problem isn’t AI — it’s context collapse

It’s 11 PM. You have 7 tabs open. Each one is an AI agent. Each one is brilliant in isolation. And each one has completely forgotten what the others are doing. Tab 1 refactored the auth module an hour ago. Tab 4 is still importing the old interface. Tab 6 is writing tests against a schema that Tab 2 changed twenty minutes ago. You’ve re-explained the project structure three times — once per agent, because none of them talk to each other. You’re not coding anymore. You’re a human clipboard. Copy context from Tab 1, paste into Tab 4. Realize Tab 4 hallucinated an import path because it never saw the refactor. Fix it yourself. Open Tab 7 to deal with the merge conflict that Tab 3 and Tab 5 created by touching the same file simultaneously, each unaware the other existed. This is what exhaustion feels like in 2026. Not “AI doesn’t work” — AI works fine. The problem is that you’re the orchestration layer, and you’re terrible at it. Not because you’re bad at your job, but because holding seven parallel execution contexts in your head while also doing the actual thinking work is a job for a system, not a person. This is context collapse. Not a failure of any single agent, but the absence of anything holding them together. Seven brilliant musicians, no conductor, no sheet music, no shared key signature. The music is cacophony — and you’re the one with the headache.

​
Genie is the conductor

Here’s the deal: you describe the work once and Genie handles the rest. It doesn’t replace your agents — it makes them actually work together. You capture your intent once. Genie broadcasts it to every agent that needs it, in a structured format they can all read. Then it runs the work in parallel — isolated worktrees, shared state, cross-agent messaging — so agents know what each other are doing without you playing telephone.
  • Capture once — write a wish, define the scope, set acceptance criteria. Done. You never re-explain.
  • Broadcast to all — every agent reads from the same plan, the same context, the same ground truth
  • Coordinated execution — parallel work across isolated branches, no merge conflicts during development
  • Agents talk to agents — Tab 4 knows what Tab 1 did, without you copying anything
You describe what you want. Genie breaks it into parallelizable work, assigns agents, manages state, and assembles a pull request. You review and merge. That’s it. 
/brainstorm → /wish → /work → /review → ship

​
What this gives you

The practical version, for the exhausted developer at 11 PM:
  • Stop re-explaining — context is structural, not conversational. Write it once, every agent inherits it. We’re serious — once.
  • Parallel execution — five agents working simultaneously across isolated worktrees. No stepping on each other’s code.
  • Reproducible workflows — wishes are versioned artifacts. Run the same plan twice, get consistent results. No more “works on my agent.”
  • Overnight mode — dispatch a team before bed. Wake up to a pull request. This is not a metaphor.
  • Bring your own agent — Claude, Codex, Open Claw, or that custom thing you built last weekend. We don’t care which AI you use.
  • Portable context — wishes, skills, and team configs live in your repo. Switch machines, switch agents, keep everything.

​
Bring your own agent

Genie has opinions about orchestration. It has zero opinions about which AI you use. It’s a protocol, not a vendor lock-in. Use whatever agent fits the job — or mix them on the same team.
AgentSetupBest For
Claude CodeWorks out of the boxDeepest integration — native plugin, team support, hooks
Codexgenie setup --codexOpenAI’s agent — full transcript and team support
Open Clawgenie setup --open-clawOpen-source option, community-maintained adapter
Custom (BYOA)genie setup --customAny CLI agent with stdout/stderr — bring your own
# Mix providers on the same team — fast engineers, thorough reviewer
genie spawn engineer --provider claude --model sonnet
genie spawn engineer --provider codex --model codex-mini
genie spawn reviewer --provider claude --model opus
Your agents are interchangeable. Switch mid-project, mix on a team, or write your own adapter. Genie cares that the work ships — not who does it.
See the Hacks page for community patterns on provider switching, cost optimization, and BYOA configurations.

​
What Genie actually does

CapabilityThe PointWhy It Matters
WishesStructured plans — scope, criteria, execution groupsYou describe it once. Agents figure out the rest.
Skills14+ slash commands — composable prompt programs/brainstorm through /dream, plus /pm for project management.
TeamsMulti-agent squads with shared stateIsolated worktrees per team. Zero conflicts during dev.
PostgreSQL BackendEmbedded database for all stateTasks, boards, agents, sessions, audit events — all in PG, not JSON files.
Boards & PipelinesKanban-style task managementConfigurable columns with gates and action skills. Project-scoped.
ObservabilityOTel-structured event loggingCost breakdown, tool analytics, session replay, error patterns — all from the CLI.
BYOAProvider-agnostic orchestrationClaude, Codex, Open Claw, your custom thing. We don’t care.
HooksEvent-driven automation on every tool callAuto-spawn, identity injection, NATS. Things happen when things happen.

​
How it works

Every feature follows the same lifecycle — vague idea to merged pull request, no detours:
1

Brainstorm

Explore the idea. Identify constraints, scope, and design decisions. The /brainstorm skill guides you through this interactively until the idea is crisp enough to plan.
2

Wish

Convert the brainstorm into a structured plan — WISH.md — with scope, acceptance criteria, and parallelizable execution groups. This is the contract between you and the agents.
3

Work

Genie dispatches agents per execution group. Independent groups run in parallel across isolated worktrees. Each agent knows its scope, its criteria, and how to validate its own output.
4

Review

A reviewer agent validates all work against the wish criteria. The verdict is SHIP or FIX-FIRST. Fix loops run automatically — up to the configured retry limit.
5

Ship

A pull request is created, targeting dev. You review and merge. The agents did the work; you made the decision.

​
Desktop is coming

Genie is CLI-first today — built for developers who live in the terminal. But we’re building a desktop app for Q2 2026: wish boards, agent dashboards, real-time monitoring. Same system, visual surface. The CLI isn’t going anywhere. If you want to orchestrate agents at 2 AM in a tmux session while your cat judges you, we respect that.

​
Get started

Features

Full feature matrix — what Genie can do today and what’s coming next.

Quickstart

From zero to your first PR in five minutes.

Core Concepts

Understand wishes, agents, teams, and skills.

Genie Hacks

Community patterns, techniques, and cost tips from power users.

Discord

Community, feedback, and support.