Quadratique Delegate issues to coding Agents
Browse documentation

Quadratique Docs

Agent execution overview

Understand the Machine to Runtime to Agent to Run execution model.

Quadratique can delegate an issue to a coding Agent running on a development machine. The Agent does the implementation work, while a member remains in control of review and merge.

The execution model

Quadratique keeps four concepts separate:

Machine → Runtime → Agent → Run

Machine

An enrolled development machine that provides compute, a repository checkout, local tools, and network access. A machine is private to its owner unless the owner explicitly shares it with the workspace.

Runtime

A coding harness detected on a machine, together with the models it makes available. Codex CLI and Claude Code are planned Runtime implementations.

The user’s local Runtime credentials and environment secrets stay on the machine. Quadratique does not store them in the Agent configuration.

Agent

A persistent, named resource that selects one Runtime and model and adds:

  • reusable instructions;
  • one or more allowed boards;
  • an owner and sharing policy; and
  • a revision number.

An Agent is not a machine or a model picker saved on an issue. Issues store the Agent selection, and Quadratique derives the execution details from it.

Run

One execution of an Agent for one issue. A run receives an immutable snapshot of the Agent’s revision, machine, Runtime, model, instructions, board permission, prompt, and repository reference.

Editing an Agent affects future runs. It does not silently change a run that is already active.

End-to-end workflow

  1. Connect a development machine.
  2. Confirm that the CLI detects a supported Runtime and model.
  3. Create an Agent for one or more boards.
  4. Delegate a To Do issue to that Agent.
  5. The machine claims the queued run and acknowledges its start.
  6. The issue moves to In Progress and exposes run details and transcript.
  7. Successful work reaches Human Review with a pull request.
  8. A member approves or rejects the work with feedback.
  9. A successful merge moves the issue to Done.

Human control points

Quadratique does not merge an Agent’s result immediately after generation. The member can inspect the pull request and run context before approving it.

Rejecting work requires feedback and creates a new run. Cancellation, interruption, and runtime failures remain visible rather than silently launching a substitute or duplicate execution.

Private by default

Machines and Agents are owner-private by default. To let another member invoke an Agent:

  • share the Agent with the workspace;
  • share its underlying machine; and
  • include the issue’s board in the Agent’s allowlist.