OmoiOS

At a Glance:

OmoiOS is an open-source orchestration runtime that turns specifications into pull requests by running parallel AI agent swarms inside isolated sandboxes, with dependency-aware execution, active supervision, and automated merging.

Overview:

OmoiOS is an AI agent orchestration runtime that reads an existing codebase, generates grounded specifications, and executes multi-agent development tasks through a structured workflow. It builds a dependency graph of tasks, runs them in parallel across isolated sandbox containers, and handles automated merging of the resulting code branches. The system moves beyond simple prompt chaining by providing a supervised execution environment where a guardian agent monitors each agent's progress, a conductor service coordinates across agents, and agents can discover new work during execution. The project includes a built-in code assistant that can explore the repository and answer questions about the codebase. It is intended for developers who need to coordinate multiple AI coding agents to work simultaneously on a real codebase with dependency awareness.

Key Decision Points:

• DAG-based execution model: Tasks are organized as a directed acyclic graph with explicit dependencies managed by DependencyGraphService, so no task executes before its dependencies are met.

• Isolated sandbox per agent: Each agent runs in its own container via Daytona, with a dedicated Git branch and filesystem, preventing interference between parallel agents.

• Active agent supervision: An IntelligentGuardian analyzes agent trajectories every 60 seconds and can inject steering interventions mid-task, rather than using a fire-and-forget approach.

• Automated merging and conflict resolution: ConvergenceMergeService merges agent branches in an optimal order, using Claude to resolve conflicts.

• Gated human approval: Users review work at strategic phase transitions and pull requests, while autonomous execution continues between gates.

Core Features:

• Codebase-grounded spec generation: The SpecStateMachine runs through Explore, Requirements, Design, and Tasks phases, generating specifications based on the actual repository structure, patterns, and dependencies.

• Parallel agent swarm execution: Multiple agents work simultaneously within their own sandboxes, coordinated through the dependency graph and critical path analysis.

• Agent supervision and steering: IntelligentGuardian scores agent alignment, detects drift, and injects steering interventions, while ConductorService monitors for duplicate work and system-wide coherence.

• Automatic task discovery: DiscoveryService allows agents to spawn new tasks when they encounter bugs, missing requirements, or optimization opportunities during execution.

• Convergence merging: ConvergenceMergeService merges sandbox branches in a calculated order, using Claude to handle conflicts.

• Built-in codebase assistant: A code assistant explores the repository structure, reads source code, and provides grounded answers, capable of spawning its own sandboxes for code generation.

Use Cases:

• Developers orchestrating multi-agent coding sessions: Run a swarm of AI agents to implement a feature across multiple files, with automatic dependency management and sandboxed parallel execution.

• Codebase exploration and grounded Q&A: Use the built-in assistant to ask questions about a repository and receive answers backed by the actual code structure and content.

• Spec-driven development from existing code: Generate specifications directly from a real codebase rather than from a blank prompt, creating a task plan grounded in the current state of the project.

Open-Source Alternative Value:

OmoiOS provides an open-source runtime for orchestrating AI coding agents, with the entire orchestration logic, supervision system, and merge service running as accessible source code. The architecture relies on structured phases, dependency graphs, and containerized isolation, allowing developers to inspect how agent tasks are planned, executed, and merged. By using Daytona sandboxes and Postgres-backed state management, the system can be run on a developer's own infrastructure, keeping the codebase and agent sessions under local control.