Gambit

Gambit is an open-source agent harness framework by Bolt Foundry for building, running, and verifying LLM workflows. It introduces the concept of "decks" — small, typed units with explicit inputs, outputs, and guardrails — that compose together into reliable AI pipelines. Decks can mix LLM calls and plain compute tasks interchangeably, feeding models only the context they need at each step instead of dumping entire documents into a single prompt.

Gambit runs locally via CLI (npx or Deno) and ships with a built-in Debug UI for streaming traces, inspecting transcripts, and reproducing failures without relying on provider logs. The framework supports REPL mode for interactive development, test-bot personas for automated testing, and session grading for evaluation.

• Typed Decks - Define workflows as small composable units with Zod-validated input and output schemas, written in Markdown or TypeScript.
• Local-First Debugging - Stream traces, inspect transcript lanes, and reproduce failures using the built-in simulator UI at localhost.
• Mixed LLM and Compute - Combine model-powered steps and pure TypeScript compute steps within the same deck tree.
• CLI Tooling - Run decks, start REPLs, launch the debug server, and grade sessions from the command line via npx or Deno.
• Test Bot and Grading - Run persona decks against your workflows and grade saved sessions to evaluate LLM behavior systematically.
• Action Decks - Attach child decks as tool actions that LLM-powered parent decks can call, enabling agent-style tool use.
• Schema-Driven UI - The Debug UI auto-renders input forms from deck schemas with defaults and a raw JSON tab.
• Session Persistence - Save and replay sessions with trace files and state snapshots stored locally under .gambit/.