AZMX AI · Solutions
Sub-agents with bounded tools and predictable outputs.
The generalist agent is great for exploration. For repeatable jobs — code review, refactor sweeps, dependency audits, incident triage — you want a sub-agent: bounded toolset, structured output, fewer surprises. Build them in YAML; ship them in your workspace; trust them once and run them forever.
Why sub-agents
When you want "the same thing, every time."
The generalist agent's strength is range. Its weakness is variance. For jobs where you want low variance — "review every PR for X," "audit every dependency upgrade," "triage every CI failure" — a sub-agent with a bounded toolset gives you the predictability without losing the AI productivity.
How to author
Four steps. One YAML file.
- Define the agent in YAMLName, description, model, tools, prompt. About 20 lines for a useful agent. Drop it in /agents in your workspace, or contribute it to the bundled catalog.
- Bound the toolsetList the MCP tools, shell commands, and read scopes the agent can use. The approval gate enforces it; the agent literally cannot reach for anything else.
- Specify structured outputWant a JSON object back? A Markdown table? A PR description? Specify the schema in the prompt; AZMX surfaces the result accordingly.
- Trust once, run foreverFirst run prompts for trust against the agent's integrity hash. After that, it's a /agent.
slash-command away in every session.
Example sub-agents
Six bundled. Each its own verb.
The generalist agent is what we reach for first. The sub-agents are what we reach for on the tenth time. Predictable output is worth a lot when you're running it a thousand times a quarter.
Staff Engineer · Series-C SaaS
Authored once, trusted thereafter.
Eighty-plus bundled skills the agent can load on demand. The sub-agent catalog is open-source — contributions welcome.