This Coding Agent Knows Which AI Model to Use for Each Task. That's the Product.

Most AI coding tools give you one model and ask you to figure out when it is and is not the right tool. Oh My OpenAgent, the multi-model orchestration plugin for OpenCode, takes the opposite approach: it routes each task to whichever model is best suited for it, and does the routing automatically.

The result, according to community benchmarks, is better output on complex work at lower total cost. Forty-eight thousand GitHub stars and 1.6 million downloads later, the routing idea appears to be landing.

How the routing works

OMO — originally oh-my-opencode, now oh-my-openagent after a rebrand that preserved backward compatibility — adds a three-layer orchestration system on top of OpenCode. The planning layer uses Prometheus and Metis agents to break down complex tasks and assign work. The orchestration layer, Atlas, routes subtasks to worker agents and manages their lifecycles. The execution layer has nine or more specialized workers: Sisyphus-Junior for general coding, Oracle for research, Librarian for documentation, Frontend for UI work, and others.

Each worker runs with model-specific prompt tuning. The default routing sends planning tasks to Claude Opus — the most capable model for architectural reasoning. Execution tasks go to cheaper models like Claude Sonnet or GPT variants. Frontend work routes to Gemini. Research tasks to Grok, which has real-time X data access. The configuration is a few JSON lines in an opencode config file.

Community testing shows OMO handling complex multi-module refactors meaningfully faster than a single-agent OpenCode setup, because workers run in parallel rather than sequentially. And because the expensive model only handles the expensive work, the token cost for equivalent output drops.

Why model routing is becoming a real product category

In 2024, the question was which AI model to use. In 2026, the question is increasingly which model to use for which subtask within a single workflow. The gap between frontier model pricing tiers has widened enough that routing matters economically. Claude Opus is significantly more expensive per token than Claude Haiku. Running every subtask through the most capable model when some subtasks do not need it is leaving money on the table.

OMO is not the only project thinking about this — the Agentic AI Foundation formed under the Linux Foundation in late 2025 is explicitly standardizing multi-agent coordination primitives. But OMO is the most widely adopted practical implementation for coding workflows, with 160 contributors and enough production usage to validate that the routing actually works on real engineering tasks rather than benchmarks.

What to watch

Ultrawork mode — full autonomy where OMO runs without human confirmation steps and corrects its own errors in a loop — is gaining traction for overnight engineering tasks. The safety conversation around unattended autonomous execution is worth tracking. Microsoft released an Agent Governance Toolkit on April 3 specifically to address this class of concern; whether governance tooling integrates with OMO’s ultrawork mode will determine whether enterprise teams can use it without the risk of unreviewed autonomous changes.

The OpenClaw migration context matters here too. OMO is a plugin for OpenCode, not OpenClaw. As developers migrate away from OpenClaw and toward OpenCode-based workflows, OMO’s user base grows as a side effect of that broader ecosystem shift.