HomerOS
What it is
HomerOS takes a simple prompt ("build a blockchain", "build a game engine", "build a compiler") and turns it into a fully orchestrated multi-agent workflow. It generates specialized agents, breaks the project into micro-tasks with a dependency graph, and runs them to completion. No configuration required.
It handles agent creation, task decomposition, model selection, inter-agent communication, memory, fault recovery, and cost control. Agents build, test, and iterate for as long as the project takes. Hours, days, or longer. No human intervention required.
How it works
The planner researches your prompt, generates a full task dependency graph, then runs a second review pass to catch circular dependencies, missing assignments, and invalid phases before a single agent fires. Agent count scales to the project. Add or remove agents at runtime.
Agents that don't depend on each other run simultaneously. A supervisor agent can delegate to workers via parsed mentions. Or run them sequentially, one turn at a time. Swap the scheduler without touching your loop.
Full state is saved after every turn: the plan, agent registry, message bus, conversation summaries. Crash at turn 200, resume from turn 200. Plug in a filesystem, memory, or custom backend.
Register tools with a schema and handler. Agents call them inline. A typed message bus lets agents address each other directly. Layered memory keeps decisions made at turn 10 accessible at turn 500. Every output is run through a validation pipeline before it lands.
Per-model circuit breakers cut failing models and route around them. Exponential backoff with jitter on retries. Multi-signal stuck detection. Three-level escalation: force output, reassign the task, skip it. A meta-agent (Moe) runs periodic reviews and can rewrite prompts, patch files, and spawn new agents.
Every LLM call is tracked: model, tokens, latency, estimated cost. Set a hard spending ceiling and the framework stops before it blows the budget. Twelve typed lifecycle events let you hook into anything. Dry-run with a mock provider and get a cost estimate before spending a cent.
Prompt to project
One function call turns a description into a complete project plan. HomerOS generates the agents, the technical specification, the system prompt, and the full task dependency graph from your input alone.
import { planProject } from './planner';
const plan = await planProject(
"Build a 2D platformer game engine in Rust"
);
// plan.agents → specialists sized to the project
// plan.tasks → 5-100 micro-tasks across phases, with deps
// plan.systemPrompt → domain-specific instructions for every agent
// plan.specContent → full technical spec with interfacesFrom there, the orchestrator runs the loop. Each turn it checks the dependency graph for actionable work, scores task complexity to pick the right model tier, and tracks completions automatically. You can swap in parallel or hierarchical execution without changing anything else.
The demonstration
To prove this works, we pointed four agents at building a cryptocurrency with a working blockchain from scratch. Proof-of-work consensus, UTXO transactions, peer-to-peer gossip, and a wallet. All in TypeScript, zero external dependencies in the core.
The agents (Bart, Marge, Lisa, and Maggie) are building it in realtime on the main page. A new commit lands approximately every minute. You can watch the block structure take shape, the UTXO engine get wired up, and the agents argue about locking scripts and mempool eviction. Live.
If the framework can handle a working blockchain, it can handle your project.