Omid Karami has released Kraft, an open-source software factory that automates the Claude Code workflow from specification to pull request. The tool addresses the persistent friction of manual agent supervision, where developers spend more time acting as schedulers and test runners than reviewing actual code. By running Claude Code in isolated git worktrees and enforcing external verification steps, Kraft shifts the human role from constant interruption to strategic decision-making.

The External Verification Gate

The core innovation is not the coding agent itself, but the separation of duties between generation and validation. Kraft executes tests in a step the agent does not control, preventing the common failure mode where an LLM claims tests pass without independent confirmation. If tests fail, a separate repair step attempts fixes, with a judge evaluating whether additional rounds are worthwhile. In a recent example, a documentation home page rebuild required two repair rounds but resulted in a clean 3,044-line addition across 30 files.

Cost Caps and Human-in-the-Loop Controls

Kraft operates on strict economic and procedural guardrails. Default spending caps are set at $10 per work item and $50 per day per instance, with the system checking limits before each agent step. Crucially, Kraft refuses to let the agent approve its own merge; it pauses for human review of the diff before opening a pull request. The system carries approvals through clean rebases, canceling merges if unauthorized changes appear. This ensures that while the agent writes the code, the developer retains final authority on what enters the main branch.

When to Skip the Factory

Karami emphasizes that Kraft is not a replacement for interactive sessions. Quick fixes, exploratory coding, and work requiring frequent steering remain better suited to direct Claude Code sessions. The ideal workflow involves thinking and planning in an interactive session, then filing the settled spec and plan with Kraft for autonomous execution. This hybrid approach allows developers to batch their attention, handling multiple work items on different chains simultaneously without context-switching penalties.

Key Takeaways

  • Kraft separates agent generation from test validation, ensuring independent verification of code quality.
  • Human approval is mandatory for merges, preventing agents from self-authorizing production changes.
  • The tool runs locally on macOS or Linux, pausing when the laptop closes to maintain security boundaries.
  • Cost caps of $10 per item and $50 per day prevent runaway API usage during autonomous runs.

The Bottom Line

Kraft proves that the bottleneck in AI-assisted development is no longer code generation, but human attention management; by enforcing external validation and strict cost caps, it transforms Claude Code from a collaborative partner into a reliable, autonomous worker.

Sources

https://dev.to/justcallmeomid/i-stopped-babysitting-claude-code-spec-in-pull-request-out-49f7