AI News

Canary is an early product from Aakash and Viswesh that provides independent verification for AI-generated or AI-assisted code changes. Claude or Codex can call Canary with a changeset, its intended behavior, and team knowledge. Canary takes a cold snapshot of the repository and can combine that information with requirements, decisions, and prior issues from tools such as Notion and Linear; it can route ambiguous questions back through the coding agent. Its harness coordinates swarms of agents from different model families, compares code before and after a change, and traces effects through callers, dependencies, state transitions, and other runtime states. The system begins with expected behavior and constraints, including keeping private files private, preventing credential leaks, and avoiding unintended duplicate effects from retries. It turns suspected failures into concrete scenarios involving actors, states, triggers, and outcomes. For each scenario, Canary selects evidence-gathering methods such as runtime verification, static analysis, unit tests, integration tests, or combinations of them. Agents execute the checks in remote sandboxes with seeded data, configured permissions, mocked dependencies, and third-party integrations. Findings and supporting evidence are returned to coding agents, which can fix the issue and request reverification. The team presents Canary as a verification harness built on top of general-purpose AI and is seeking early user feedback. A CLI is available through the `@runcanary/cli` npm package.