Models
The model is one field in the agent's YAML. Run Claude models on Claude Code, OpenAI models on Codex, or open-weight models on the same harness, and route tokens through your own Anthropic key or AWS Bedrock account when you need to.
Every selectable model is listed with the id you name in YAML and its published token rates. Put the reviewer on a frontier model and the digest agent on a cheaper open-weight one, in the same repo.
Every selectable model, with the id you name in YAML
Switching models is a one-line edit to the config. The sandbox, credentials, budget, and transcript stay the same, so you can try a cheaper model on one agent and compare the cost breakdown before rolling it out.
session: harness: type: claude_code model: claude-opus-5 # Or replace that harness with: # {type: claude_code, model: glm-5.2} # {type: codex, model: gpt-5.6-terra}claude.model and codex.model, one line each
Connect an Anthropic key or a Bedrock role and token spend lands on your account, against credits and commitments you already hold. Budgets and cost breakdowns work the same either way, and the provider credential never enters a sandbox.
llm: proxy: proxycfg_8f2k1x9qYour key or your AWS account, verified before it routes