Zeroshot

Zeroshot turns a software goal into an explicit multi-agent graph: one agent implements the change, independent agents review it, rejected work goes to a bounded repair loop, and nothing is delivered until the graph's checks pass. The implementing agent never approves its own work. Zeroshot does not replace a coding agent. It runs Claude Code, Codex, or GitHub Copilot as both the worker and the reviewers, and local runs can reuse that harness's existing login.

In the built-in software-change graph, a worker implements the task, then acceptance and code reviewers check the result in parallel. Rejected work goes to one repair worker with both reviews' feedback and back through both reviews. With delivery enabled, accepted work continues through the configured Git and CI steps, and delivery conflicts return through repair and review. The graph defines the sequence, parallel steps, retry paths, and exit conditions before execution starts, and every run records its events in a durable SQLite ledger.

A passing run means its configured checks accepted the work, so coverage depends on the requirements, reviewers, tests, and environment you supply. You can extend the graph with your own stages (adversarial tests, end-to-end tests) and save a graph with its per-node model settings as a reusable profile, shared between the CLI and the browser UI.

Zeroshot ships as a native executable. Graphs run locally or on a self-hosted Docker target that bundles the engine with the Codex, Claude, and GitHub Copilot harness CLIs and serves its own profile editor and run viewer.