Zeroshot: coding agents with independent review and bounded repair

Zeroshot coordinates coding agents through a defined implementation-and-review process. Its official repository describes one agent writing the change while separate agents assess it, with rejected work returning to a bounded repair loop. The implementing agent does not approve its own work.

Separate implementation from acceptance

In the built-in software-change graph, acceptance and code reviewers run in parallel. If either rejects the result, a repair worker receives their feedback and both reviewers check the revised work again. Git delivery can follow acceptance, with delivery conflicts also returning through repair and review.

The repository states the limit of that process: a passing run means its configured checks accepted the work. Coverage still depends on the requirements, reviewers, tests and execution environment. A successful graph is therefore evidence about a particular review setup, not proof that the resulting software has no defects.

Zeroshot runs installed coding harnesses as its workers and reviewers. Claude Code, for example, supplies codebase reading, file edits and command execution; Zeroshot describes the graph that assigns and checks that work. This distinction helps a developer evaluate the review policy separately from the chosen coding agent.

Define the graph and keep a durable run record

Users can start with the built-in graph or define a topology with their own parallel steps and retry paths. The repository says custom stages such as adversarial tests or end-to-end tests require the user to define their tools, services and credentials. Those stages are not all supplied by the default review loop.

Profiles save a graph and runtime settings for reuse. The CLI and browser interface share local profiles, while a durable SQLite ledger records run events. Inspect those events and the exact reviewer feedback when deciding whether an accepted change is ready for your repository.

Choose the execution and delivery scope

The installation instructions require Node.js 18 or newer and describe a native binary for supported Linux, macOS and Windows architectures. Local execution also requires an installed, signed-in supported harness such as Codex, Claude Code or GitHub Copilot.

The repository's example keeps delivery local and warns that the worker edits the current Git worktree. Begin in a clean worktree dedicated to the task and validate the graph, input and runtime configuration before execution. Optional delivery modes include pushing a managed branch, preparing a pull request or proceeding through pull request, CI and merge.

Self-hosted targets and Zeroshot Cloud offer other execution arrangements, each with its own tools and authentication. Franklin has not installed Zeroshot or run a software task; this entry explains the repository's documented behavior rather than a hands-on result.