3 min read

[Series] Harbor Evals for Agentic Skills

Cursor skills grow fast. The hard part is not writing the first SKILL.md - it is knowing the skill still does what you need after the next edit, the next model, or the next teammate’s prompt.

This series is about Harbor evals for agentic skills: repeatable checks for agent behaviour, the Harbor (opens in a new tab) CLI that runs those checks in containers, and Eval-Driven Development (EDD) so you can improve a skill against an eval instead of only against manual eyeballing.

If you followed Building Scalable Skills for Agentic Workflows (opens in a new tab) to build a Team Brain, that skill’s Intent identified: line is the contract many of the sample evals assert.

Who Harbor evals for agentic skills are for

  • People authoring Cursor skills (or similar agent workflows) who want regression checks
  • Engineers who know Test-Driven Development (TDD) and want the same rhythm for agents
  • Anyone willing to pay their own model API costs for containerized trials - Harbor orchestrates; you bring the keys

What you will build toward

Companion samples live in CodeSloth Cursor Samples (opens in a new tab):

Posts in this series

Only published parts are linked. Draft parts stay named until they ship.

PartTopic
Part 1 - What Are Evals? (opens in a new tab)Why the manual skill loop hurts, EDD vs TDD, what Harbor is
Part 2 - Install and Scaffold (opens in a new tab)Install Harbor with uv, harbor tasks init, default task layout
Part 3 - Writing Tasks (opens in a new tab)Pass/fail fixtures, deterministic checks vs LLM-as-judge, expected output files
Part 4 - Task Shapes and DRY (opens in a new tab)Granularity, base images, mounts, generators, orchestration tradeoffs
Part 5 - Running and ResultsRun scripts, API keys, model matrices, harbor view
Part 6 - Eval-Driven DevelopmentInterview → write eval → improve Team Brain in a capped loop

How to read Harbor evals for agentic skills

Start with Part 1 (opens in a new tab) if you want the mindset. Skip to Part 2 (opens in a new tab) if you already buy the idea and want the CLI on your machine. Jump to Part 3 (opens in a new tab) when you are ready to write pass/fail fixtures and verifiers.

Official Harbor docs: harborframework.com/docs (opens in a new tab). Harbor source: github.com/laude-institute/harbor (opens in a new tab).

Sloth Summary

  • Harbor evals for agentic skills turns skill quality from “looks fine in chat” into Harbor tasks with verifiers and rewards.
  • Parts 1-4 are published; Parts 5-6 follow.
  • Samples grow under harbor-evals/ and team-brain-edd/ in CodeSloth Cursor Samples.

May your regressions show up in a reward file, not in a teammate’s chat. 🦥

Keep reading

More in AI & Sloth Series

Harbor Task Shapes and DRY (Harbor Evals Part 4)

Harbor task shapes and DRY patterns: simple skill evals vs complex flows, shallow-clone base images, sibling tasks, skill volume mounts, and generate/copy scripts.

10 min

Harbor Evals Part 3: Writing Tasks with Pass and Fail Fixtures

Writing Harbor tasks with reference pass and fail outputs, deterministic pytest checks versus LLM-as-judge, neutral instructions, and output files the verifier can find.

12 min

Harbor Evals Part 2: Install Harbor and Scaffold Tasks

Install Harbor with uv, explore harbor --help, and scaffold tasks with harbor tasks init - plus the default folder layout agents can also create for you.

6 min