Agents, models, and options

Choose a Runtime harness, pin the model, and record operational options safely.

On this page

Built-in harnesses

Kind	Credential variable	Typical use
Claude	`ANTHROPIC_API_KEY`	Claude Code CLI harness.
Codex	`OPENAI_API_KEY`	Codex CLI harness.
Antigravity	`GEMINI_API_KEY`	Google Antigravity CLI harness.
Muse	`META_API_KEY`	Meta endpoint with optional checkpoint and base URL.
Grok	`XAI_API_KEY`	xAI CLI harness.
OpenCode	Provider-dependent	`OPENCODE_API_KEY` or `OPENROUTER_API_KEY`.
Terminus	Provider-dependent	`ANTHROPIC_API_KEY` or `OPENAI_API_KEY`, selected by the provider-qualified model.
DeepSeek	`DEEPSEEK_API_KEY`	A `deepseek/…` provider-qualified model.
Scripted / Python	None declared by the harness	A pinned script or Python factory; Task or Verifier bindings may still apply.

Pin models explicitly when reproducibility needs it

The Registry composer treats model as a form field and records it in the agent options; Muse may be left blank. The Runtime CLI has no separate model flag: use --agent-option model=PROVIDER/MODEL when the harness supports it. Dynamic provider-qualified harness defaults are canonicalized, while an omitted static-harness model may remain that harness's implicit default. UI suggestions are conveniences, not a universal allowlist.

Agent options

Terminal

fleet-runtime runs submit benchmark -n 1 \
  --agent claude \
  --agent-option max_turns=20 \
  --agent-option max_budget_usd=5 \
  --harness-var ANTHROPIC_API_KEY=@ --follow

Option values are recorded in the durable Run spec. Do not put credentials in options or a URL; use a declared harness variable or Registry credential source.