Search Documentation

Find and navigate to documentation pages and their content

Agents & Subagents

Codistry can hand a self-contained chunk of work to a subagent (a fresh, isolated assistant with its own brief and its own tools) and fold the result back into your conversation. It keeps the main thread focused and lets you scope risky work tightly.

On by default

Subagents are enabled out of the box (adronite-agent.subagentsEnabled defaults to true). The main agent decides when to delegate, and you can nudge it too: “use a subagent to research how auth works.”

Why delegate to a subagent

Keep the main thread clean

Research and verification run in isolation: only the final report comes back, not all the intermediate steps.

Scope tools tightly

Give an agent a narrower set of tools so it can only do what its task needs.

The built-in agent

A general-purposeagent is always available, so delegation works without any setup. It's well suited to searching across a large codebase, reading many files to understand how something fits together, and other self-contained, multi-step investigations.

Create your own with AI

Open Settings → Agents and describe what the agent should do and when to use it. Codistry writes the definition with your configured model, saves it, and opens it for review. You can edit it anytime.

Good agents have a clear job

Describe a focused role and when to use it, for example, “a test-runner agent that runs the suite, reads failures carefully, and reports the most likely cause without modifying files.”

Where agents live

TierLocationNotes
Built-inShips with CodistryThe general-purpose agent. Always available.
Your agents~/.adronite/agents/Created and edited by you, fully trusted.
Project agents.codistry/agents/Committed with a repo and shared with your team.

What an agent can do

From the Agents tab you can shape each of your own agents:

Allowed & disallowed tools

Limit an agent to a subset of tools, or block specific ones. By default an agent inherits the same tools the main agent has.

Auto-approved commands & max turns

Optionally let an agent run certain command prefixes without prompting, and cap how many steps it may take per run.

Trust & safety

Delegating never bypasses your approvals
  • A subagent's tool actions still ask for your approval, labeled with the agent's name, the same prompts you'd see normally.
  • Project agents from a cloned repo can narrow tools but can never auto-approve commands on your machine.
  • Subagents can't spawn further subagents.

Want to turn the feature off entirely? Disable subagentsEnabled.

Related