opencode-agent-creator

To create reusable OpenCode agent profiles, define specialized primary agents or subagents using Markdown files or JSONC entries with controlled prompts, models, tools, and permission approval flows. You configure capabilities like temperature, maxSteps, and tool availability to establish operational guardrails.

The best way to manage subagent permissions is by configuring permission approval flows to allow, ask, or deny specific risky actions like edit, write, or bash commands. You can also control which subagents can be invoked through task permissions to limit operational impact.

Set up a restricted code-review subagent by creating a Markdown profile under the .opencode/agents/ directory. Configure the agent with a description, appropriate mode, system prompt, and tool permissions to allow safe Git inspections while requiring confirmation for potentially risky commands.

Yes, you can configure temperature, maxSteps, and model selection for agents via opencode.jsonc or Markdown profiles. The configuration schema supports fields including name, description, mode, model, tools, permission, hidden, disable, and optional prompt loading from inline text or file paths.

OpenCode agent configuration supports loading system prompts from external files using the {file:path} syntax. This allows you to separate prompt logic from configuration metadata when defining agent profiles via Markdown or JSONC.

Markdown profiles are recommended for creating OpenCode agents to ensure consistent prompts and tool access across primary agents and subagents. JSONC configuration via opencode.jsonc is also supported, but Markdown provides a structured format for reusable behavior definitions.