Open subagent definitions

generated, not hand-written

The subagentjobs monorepo does not hand-edit its Claude Code sub-agent files. Every YAML under .claude/agents/ is generated from one typed Rust source of truth — so an agent's model, tools, and prompt are compiler-checked data, not free-form markdown that drifts.

The 10 generated agents

Baked 2026-07-09 from the real generated files: grep -E '^(name|model):' .claude/agents/*.yaml. Static page — regenerate the agents and this table drifts until the next authoring pass; the live, queryable catalog is subagentcode.com.

The generation chain

1. crates/schema/src/agent.rs — typed AgentConfig / AgentModel / AgentTool enums with validate() and to_yaml(). A model id that isn't an enum variant does not compile.
2. cargo run -p agent-gen — crates/agent-gen/src/agents.rs's all_agents() declares the 10 agents and regenerates every YAML.
3. .claude/agents/*.yaml — the generated output Claude Code actually loads.

One real consequence of the typing: when Claude Fable 5 access was restored (2026-07-01), adding it meant adding a Fable5 variant to the enum — and exactly one agent, engineering-coworker (the one that executes deploys, git pushes, and D1 writes on a real Mac), was upgraded to it. The other agents deliberately stayed on Haiku/Sonnet. That decision is visible in the table above.

What "open" means here

Honestly scoped: open means the definitions are public in the repo — readable source (agents.rs), readable output (.claude/agents/*.yaml), and a public JSON API over the catalog at subagentcode.com (GET /api/agent-definitions). It is not a registry accepting outside submissions, and the agents themselves run against this repo's own infrastructure.