Claude Code Harness rebuilt in Go

Claude Code Harness v4 is an open-source wrapper around Anthropic’s Claude Code that tries to turn agentic coding into a controlled delivery loop rather than a loose sequence of prompts. The project’s public repository describes a fixed path — setup, plan, work, review and release — with hooks, rules and audit artifacts designed to keep long-running coding sessions on track. Its maintainer says the latest runtime was rebuilt in Go, replacing earlier overhead with a “Go-native guardrail engine” that responds in under 10 milliseconds. Anthropic has been making a parallel argument in its own engineering posts. In November 2025, the company said effective harnesses matter because agents working across many context windows need structure borrowed from human engineering workflows. In March 2026, Anthropic said harness design is central to pushing Claude on long-running application development. ### What is this harness actually doing? (github.com) The GitHub repository describes Claude Code Harness as a “Claude Code Dedicated Development Harness” built around an autonomous Plan→Work→Review cycle. The README says the harness is meant to manage “everything else” around code generation — including plans, reviews and the paper trail needed to trust the output. The project organizes that into a release-oriented loop. (anthropic.com) The public materials say work begins with a plan, moves into parallel execution, then passes through multi-angle review before release, with a changelog and validation artifacts left behind. That makes the harness closer to workflow orchestration than to a standalone model or coding assistant. ### Why does the Go rebuild matter? The maintainer’s repository notes say the current v4 line is “Go-native,” and the public GitHub summary says the guardrail engine protects execution with “sub-10ms response.” A mirrored repository summary and social post cited in the source briefing both describe earlier overhead in the 40–60 millisecond range, though that figure appears in project materials rather than an Anthropic benchmark. (github.com) That latency claim matters because harnesses sit in the control path. (github.com) If every rule check, hook or validation step adds delay, the overhead compounds across long sessions and multi-agent runs. Anthropic’s own engineering posts do not cite this project specifically, but they do frame harness design as a performance and reliability issue for long-running agents. ### How is this different from Claude Code itself? (t.co) Anthropic’s official Claude Code repository describes Claude Code as an agentic coding tool that lives in the terminal, understands a codebase and can execute routine tasks and git workflows. Its docs publish a separate changelog for the core product, including recent infrastructure and security fixes. Claude Code Harness sits on top of that layer. (anthropic.com) The harness repository says it adds workflow contracts, policy checks, hooks, path-scoped rules, review stages and release bookkeeping. In practice, that means Claude Code remains the coding engine, while the harness acts as the operating procedure around it. ### Why are audit trails and changelogs central here? The repository says release steps are tied to changelog updates and validation reruns, and recent commits include release-completion markers and changelog edits for v4.10.0. (github.com) The project also includes templates and audit-trail components, according to the public file tree and commit history. That emphasis matches Anthropic’s broader direction for managed and long-running agents. (github.com) In an April 2026 engineering post, Anthropic said assumptions built into harnesses need to be revisited as models improve, and in its managed-agents post it said harnesses encode operational constraints around what models can and cannot do. ### What can be verified today? GitHub shows the project is public, under active development and recently updated, with v4.10.0 release-related commits posted in May 2026. (github.com) Anthropic’s engineering blog separately confirms that “harness design” has become a named area of work around Claude and long-running agents, even though this specific open-source harness appears to be a community project rather than an official Anthropic release. (anthropic.com) The next concrete place to watch is the repository’s changelog and migration documentation, which track v4 updates and release mechanics, alongside Anthropic’s Claude Code changelog for changes in the underlying tool the harness is built around. (github.com 1) (github.com 2)