This is a sample configuration for the Opencode orchestrator.
This config is highly specific to me. Use it to get the general idea of how to configure the orchestrator, but adapt it to your needs!
Mainly, make sure to set the subagent model to an extremely cheap one and for the main agent use a SOTA model. I use GPT 5.5 for the main agent and Deepseek-v4-flash for the subagents.
Basically, you use the Orchestrator main agent with a smart model (My instructions work fine for GPT 5.5).
You limit the tools it needs. Exclude MCPs that subagents can do, include ones that can't. I use the 3 subagents:
- Scout - For doc research. Uses Deepwiki, Context7 and Exa/Firecrawl. Looks up docs, if needed, references files.
- Explore - Basic read-only agent that verifies if files are ok and also maps out the codebase.
- General - Uses all the MCPs, runs the commands, edits the files. Basically the "Build" agent you'd use by default.
The configuration is a bit verbose, but that's because it severly limits what tools/mcps each agent can use. The main one is allowed to use all, but as a last resort, when the subagents spectacularly fail.
Subagents are limited to only the tools they need, and the main agent is limited to only the tools it needs. Tweak this according to your own workflow and needs.
The main agent delegates tasks to the subagents, and the subagents execute them. It does this in 5 phases.
- Domain mapping - The main agent maps out the domains (modules, layers, services) involved in the task.
- Context gathering - The main agent gathers context from the domains.
- Plan - The main agent creates a plan for the task. (You should specify if it should output a simple one for itself, or write to a file)
- Execution - The main agent spawns subagents to execute the plan, one per domain. Each subagent works independently on its assigned domain.
- Review - The main agent spawns a multitude of verification agents to review the results and ensure quality.