Mixedbread Search Agent Harness

Docs: https://www.mixedbread.com/docs/agent/build-your-own-harness API: https://www.mixedbread.com/api-reference/endpoints/chat/create-chat-completion Model: https://www.mixedbread.com/docs/agent/models Public training harness: https://github.com/mixedbread-ai/toast-harness

Toast gathers evidence through tools and returns a ranking, an answer, or both. It can work with different retrieval backends, tool sets, output formats, budgets, and agent frameworks. The public harness is a useful implementation reference; reproducing it is not a condition for good performance. The design guidance below describes preferences to evaluate on your task.

API constraints and design choices

The API's accepted schemas and fixed model behavior are constraints: use toast-1, keep thinking disabled, and stay within the context window and supported parameter ranges. The hosted model already disables thinking; chat_template_kwargs is not an API setting. When continuing a Chat Completions tool turn, supply one result per tool_call_id, including failed or rejected calls. Stored continuation requires that the preceding completion was stored.