Show HN: Gambit, an open-source agent harness for building reliable AI agents
Gambit is trying to turn “agent glue code” into something you can actually reason about. Instead of writing one huge prompt that calls a pile of tools, you build a tree of small, typed steps (“decks”) with explicit inputs/outputs and guardrails. Some decks call models, some are pure compute, and you can mix them in a single workflow.
The most practical part is the local-first debugging story. Gambit can stream traces to a built-in Debug UI, persist state under .gambit/, and includes commands for running a deck once, iterating in a REPL, and even grading sessions. If you’ve ever struggled to reproduce “it worked yesterday” agent failures from provider logs, having deterministic local traces is a big upgrade.
If you want to try it, the README suggests starting with npx @bolt-foundry/gambit init, then running an example deck via repl (terminal) or serve (browser). The project also exposes a TypeScript API (JSR import) where you can define inputSchema/outputSchema with Zod and implement compute steps alongside model calls.