OpenCode Agent Creator

Confirm the target OpenCode generation first. V1 and V2 beta configuration and plugin APIs are not interchangeable; preserve an existing V1 setup unless the user explicitly requests a V2 migration and the current V2 docs have been checked.

Build precise, functional, professional OpenCode agents and agent teams without re-researching the platform. The reference shape is one visible primary "supervisor" that plans and delegates, dispatching a roster of model-tiered subagents through the native task tool. This skill encodes the official OpenCode agent model plus field-tested orchestration, topology, and cost-routing patterns.

Scope: this skill creates OpenCode agents (the runtime actors). For repo instruction files use agent-md-creator; for portable Agent Skills use skill-creator; for Claude Code subagents use agents-claude-creator. An OpenCode subagent often loads a Skill to get its methodology — the two compose via the skill tool.

The one constraint that shapes every decision

OpenCode fixes the model per agent. The native task tool accepts only description, prompt, and subagent_type — there is no model parameter. A primary agent cannot pick a model per dispatch at runtime. Therefore:

• Cost control = routing, not runtime selection. Pre-create agents pinned to different models; the supervisor controls cost by choosing which agent to dispatch — a cheap model for trivial/bulk work, a strong model for hard reasoning, each a separate agent.
• "Create a team with the best/cheapest models" is a config-time act (writing model: into each agent), not something the supervisor does live.
• To get "cheap for easy, strong for hard" inside one domain, make two agents (e.g. x-operator and x-operator-heavy) and let the supervisor route between them.

Internalize this first — it is why the architecture looks the way it does.

The cold-context contract (the other thing that shapes every prompt)