Agents on Claude (Beta)
Agents are in Beta. An Agent is a system prompt plus the Claude access it runs on; an AgentRun is one conversation with it, working in its own sandbox.
- The right model for each prompt. On Nodus’s Claude, every prompt is routed to Haiku, Sonnet, Opus or Fable, and that model answers the whole turn, including its commands. If routing is unavailable, Sonnet answers. Each turn records the model that served it and what it cost.
- Runs survive restarts. A run that is interrupted continues where it stopped and never repeats a model call or command it already made.
- Pay as you go, or bring your key. Usage is billed from your credits at cost divided by 0.875, capped per run
by
spec.perRunMaxCostUSD. Withmodel.access: BYOKyour own Anthropic key pays instead and Nodus charges nothing for the model. - Follow-ups from the CLI.
nodus create agentrun NAME --agent claude-assistant --prompt "..."starts a run, andnodus agentrun send,stepsandanswermessage it and read what it did.
Fan-out, groups and evals are not available yet. See Agents.