Choose the coding agent

You'll learn which coding agents a conductor can launch, how to pick one per project and per ticket, and what each agent needs installed on the conductor's machine.

The agent is a surface, filled by an addon. A conductor's engine config lists one or several agent addons; at dispatch the conductor scores them for the ticket at hand and launches the winner. Nothing else in the loop knows which agent it is talking to.

The agent addons

Addon	Launches	Footprint
@gaia-ai/addon-claude	Claude Code	parsed from the real transcript
@gaia-ai/addon-codex	OpenAI Codex CLI	parsed
@gaia-ai/addon-pi	pi	parsed
@gaia-ai/addon-kimi	Kimi	stubbed — an empty footprint until the parser lands
@gaia-ai/addon-opencode	OpenCode	stubbed
@gaia-ai/addon-grok	Grok Build	ships as a conductor agent addon

The footprint is what a run reports for billing — tokens, duration, turns, tool calls, model. An agent whose footprint is stubbed still works; its runs just report zeroes.

Every agent addon launches that vendor's own CLI inside the run's tab. The CLI must be installed and authenticated on the conductor's machine — the addon starts it, it does not ship it. A conductor that can only run Claude Code lists only the Claude addon.

One agent for the project

The simplest engine config names one agent:

// .gaia/conductor.config.js
agent: { plugin: '@gaia-ai/addon-claude' },

Every run of every ticket this conductor claims uses Claude Code.

Several agents, chosen per ticket

agent may also be an array of candidates. A candidate is either a bare descriptor or a { agent, priority } wrapper, where agent is the descriptor and priority(ticket) a function. At dispatch the conductor calls every candidate's function with the ticket — enriched with its labels as a list of label names and its environments — sorts highest score first and launches the top one. Ties go to the earlier entry.

agent: [
  { agent: { plugin: '@gaia-ai/addon-claude', with: { model: 'claude-opus-4-8' } },
    priority: () => 0 },
  { agent: { plugin: '@gaia-ai/addon-codex' },
    priority: (ticket) => ticket.labels.includes('codex') ? 100 : -Infinity },
],

With that config a ticket labelled codex runs on Codex; everything else runs on Claude Code.

Only the wrapper carries a priority

A priority written beside plugin in a bare descriptor is ignored — the loader reads it only from the { agent, priority } wrapper. A candidate without a priority scores -Infinity; it wins only when every candidate scores that low, and then the first in the list wins. Give your default agent an explicit baseline such as () => 0, and let conditional candidates return a higher score when they apply.

The labels GAIA ships for this — codex, kimi, kimi-opencode, pi, grok — are just terms on the control plane; which of them your conductor honours is decided entirely by the priority functions in its engine config. See Labels & work types.

The ticket-create agent

The agent that gaia ui starts when you hand a new ticket to it is not a run and does not go through the engine config. The cockpit launches a command from your machine context, ~/.gaia/machine.config.js: the agent_command key names the binary, and it defaults to claude. Set it to another agent's CLI to create tickets with that agent instead:

// ~/.gaia/machine.config.js
export default { /* … */ agent_command: 'codex' };

Next

• Engine config — the agent slot, entry forms and options in full.
• Run a ticket: from claim to done — where the chosen agent runs.
• Labels & work types — the labels a priority rule can read.