Type Three Letters to Let an AI Team Code for You

Oh My OpenAgent (OMO) introduces the ultra‑short ulw command—standing for Ultra Work—to launch a completely hands‑free development workflow. After typing ulw, developers can step away while the system generates a full feature, achieves over 90% test coverage, produces API documentation, and automatically patches potential SQL‑injection and XSS vulnerabilities.

As of June 2026, OMO has earned 62.3 k GitHub stars, over three million downloads, and roughly 600 k monthly active developers. Users report tasks ranging from fixing thousands of ESLint warnings to building entire web applications in minutes, with some abandoning other AI assistants such as Cursor and GitHub Copilot.

ulw Mode – Seven‑Layer Execution Engine

Intent Gate: The Sisyphus coordinator parses the developer’s true intent, infers hidden requirements from repository history, and auto‑generates appropriate models (e.g., SQLAlchemy for PostgreSQL) and RESTful endpoints.

Parallel Codebase Exploration: Within 100 ms, three Explore agents scan the codebase, map architecture, and identify existing authentication, middleware, and database schemas.

Intelligence Sync & Planning: The Librarian agent fetches OAuth 2.0 specs, JWT best practices, and known vulnerabilities, feeding a Prometheus planner that emits a YAML development plan with task breakdown, dependencies, acceptance criteria, and risk assessment.

Dual‑Plan Verification: Metis checks the plan for missing edge cases (e.g., token expiry), while Momus performs adversarial review to surface security flaws; only plans passing both proceed.

Multi‑Agent Parallel Execution: Atlas, the main executor, spawns 1‑8 specialized agents—backend, frontend, testing—to work concurrently on API code, UI, and unit tests.

Closed‑Loop Self‑Correction: After code generation, automated test suites, quality checks, and security scans run; any failures trigger automatic debugging, code modification, and re‑testing until all checks pass, with a detailed correction log recorded.

Atomic State Persistence: Every atomic operation’s state is saved to boulder.json, enabling exact recovery after power loss or network failure.

A head‑to‑head experiment showed that building the same authentication system with Cursor required 17 human prompts and 2 h 15 min, whereas OMO’s ulw mode completed it in 28 min with higher code quality and test coverage.

Why OMO Achieves What Others Cannot

OMO’s architecture departs from single‑agent designs by mimicking a professional development team with strict role separation:

Sisyphus (CTO): Powered by Claude Opus 4.7 Max, it orchestrates overall decisions without writing code.

Prometheus (Planner): Based on Claude Sonnet 4.6, it produces structured design artifacts only.

Atlas (Executor): Also Claude Sonnet 4.6, it enforces the plan, routes tasks to the most suitable agents, tracks progress, and retries on failure.

This “plan‑execute‑verify” triad eliminates hallucinations and self‑review errors because planning mistakes are caught before execution and verification catches runtime issues.

Team Mode further scales parallelism: enabling team_mode.enabled=true allows up to eight specialized agents to collaborate via twelve dedicated team tools covering lifecycle management, task assignment, communication, status queries, quality control, and exception handling.

Dynamic Agent Assembly

OMO does not use a fixed team; it assembles agents on‑the‑fly based on task type using a “category + skill” system. Categories map to optimal base models (e.g., ultrabrain → GPT 5.5 xHigh for complex reasoning, deep → GPT 5.5 Medium for heavy debugging, visual‑engineering → Gemini 3.1 Pro for UI work, writing → Kimi K2.5 for documentation, git → Claude Haiku 4.5 for version control, quick → GPT 5.4 Mini for simple tasks). Skills such as React, TypeScript, Tailwind, FastAPI, PostgreSQL, Redis, and OWASP‑Top10 are injected into the respective agents.

Cost‑optimization is built in: OMO automatically selects the cheapest model that satisfies the task, saving roughly 75 % of API cost compared with using a single GPT 5.5 model, while maintaining comparable output quality.

Engineering Lessons for Multi‑Agent Platforms

Define clear responsibility boundaries before building agents.

Invest in shared infrastructure (unified task model, reliable message bus, distributed state store, tool abstraction layer, observability system) prior to agent development.

Ensure reliability through idempotent operations, transactional sandbox execution, tiered retry strategies, and graceful degradation.

Embed cost awareness via hierarchical model scheduling, incremental context syncing, result caching, and quota management.

Finally, the article argues that AI is shifting from an “assistant” that follows step‑by‑step instructions to an autonomous “team” that only needs high‑level goals, freeing developers to focus on problem definition, architecture, and complex business logic rather than repetitive coding.