Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    HarnessLab

    1

    Build, trace, replay, diff, benchmark, and optimize AI agent harnesses with a reproducible Go CLI.

    Free

    Security scanned
    HarnessLab

    HarnessLab

    Example session with this skill installed

    I have a recorded coding-agent run. Show me what happened, replay it offline, and tell me whether the fix actually passed.

    • Read your context and instructions
    • Compiled the harnesslab

    The agent lists the run, prints the trace, runs harness replay <run-id>, and reports verification PASS/FAIL, model/tool call counts, token/cost summary and artifact paths. Strict replay makes no model API calls.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    HarnessLab is a Go platform for AI Agent Harness Engineering. It turns prompt, context, planning, tools, verification, retry, sandbox, budget and pricing into a declarative harness.yaml that can be observed, replayed, compared and optimized.

    The CLI covers the full loop:

    • harness run: execute a coding task in an isolated Git worktree.
    • harness trace: render the normalized model/tool event stream.
    • harness replay: re-run a recorded trajectory offline without model API calls and re-apply the recorded workspace patch before verification.
    • harness diff: align two trajectories and find the first divergence.
    • harness bench: run task × harness matrices with dev/holdout splits, repeated runs, token/time budgets, deterministic verification and statistical reports.
    • harness optimize: analyze failures, generate candidate harnesses, evaluate on dev, compute a Pareto front and validate on holdout.
    • harness import-swebench: import SWE-bench Verified tasks with test_patch, FAIL_TO_PASS and PASS_TO_PASS preserved.

    It supports tRPC-Agent-Go and Codex CLI runtimes behind one runtime-agnostic interface, plus process/Docker/bubblewrap sandboxes, cost accounting, reproducibility bundles and environment fingerprint validation. A model claiming “fixed it” is not a PASS: verification and, optionally, a real workspace change must confirm the result.

    Where it stops: HarnessLab provides the harness, replay and evaluation plumbing. It does not install repository dependencies automatically, and a process sandbox is not a hard security boundary. Full SWE-bench evaluation still needs repository-specific test runners or the official SWE-bench Docker harness.

    Does not install repository dependencies automatically; process sandbox is not a hard security boundary; full SWE-bench evaluation requires repository-specific test runners or the official Docker harness.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Free forever

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 21 days ago

    • Free with an Agensi account

    Requires

    OpenAI
    GitHub
    Deepseek
    Go 1.25+
    Git
    DeepSeek or OpenAI API key
    Docker or bubblewrap

    Permissions

    Allowed Hosts

    api.deepseek.com
    api.openai.com
    github.com

    File Scopes

    .harness/
    repository
    worktree

    API keys are read from environment variables and never written by the skill.

    Listed21 days ago

    What's inside

    Frequently Asked Questions

    Strict replay does not. It serves recorded model and tool calls from the replay store. --fallback allows live calls on a miss.

    It can run any Git repository, but you must provide verification commands that are valid for that repository.

    No. A small self-authored task set is a case study. Use dev/holdout, repeated runs, an external benchmark and paired analysis before making a general claim.

    It adds cwd isolation, environment scrubbing, timeouts and command policy, but it is not a hard security boundary. Use Docker or bubblewrap for untrusted code.

    No. Use environment variables and keep .harness artifacts out of source control.

    New in LLM Ops