Tuning Your Agent
Put numbers on your agent's output and improve it deliberately. Review evals, task evals, and the optimizer. Measure quality, then move the bar.
There are two ways through this course. You can read each lesson on this site and step through it here, or run it hands-on in your own coding agent. The two cover the same ground, so pick whichever suits you, or do both.
To run the lessons in your agent, install the course once. Do this in a fresh project directory (for example a new tuning folder, because Tessl won't initialize in your home directory):
npx tessl install tessl-academy/tuning-your-agent
This installs the skills your agent uses to walk you through each lesson interactively. The course ships as a single plugin bundling a guided walkthrough for every lesson (01-reviewing-skills-with-review-evals, 02-task-evals-and-scenarios, and so on).
Once the course is installed, start a lesson by asking your agent. Open your coding agent (Claude Code, Cursor, Codex, or Tessl Agent) in that directory and ask it:
"guide me through reviewing skills with review evals"
The installed skill picks up the request and walks you through the lesson step by step. Prefer a command? Launch it directly:
tessl launch skill --agent claude-code -i 01-reviewing-skills-with-review-evals
Swap claude-code for cursor, codex, or tessl-agent. (Install must run first, because tessl launch skill does not install for you.) Each lesson hands off to the next, so your agent keeps its context as you move through the course.
You'll work on a real skill throughout. If you completed Skill Foundations you already have a commit-conventions skill and a git-hygiene plugin to tune; if not, install the foundations course (npx tessl install tessl-academy/skill-foundations) for a ready-made one.
Reviewing skills with review evals
Score a skill's structural quality with a review eval, read the report, and gate publishes on a threshold.
Task evals & scenarios
Generate behavioral scenarios from real commits, bend the rubric to what you mean by good, run a task eval, and read the baseline-vs-with-context lift.
Optimizing a skill
Run the review-eval optimizer to raise a skill's score automatically, then dogfood it against real work to catch what the optimizer can't see, before you publish.
Reviewing skills for security
Publish a skill that carries a real vulnerability, read the Snyk finding the security review surfaces on the registry, then install it and decide what to do with the gating prompt.
Continuous review in CI
Point a production CI gate at a repo with a failing skill, watch it block the PR, then optimize the skill until the gate goes green.
Comparing models for cost vs quality
Run the same task eval across several models, read the cost-vs-quality results, and pick the cheapest model that still does the job well enough.
Repo autopilot with Tessl Agent
Connect a repo to Tessl Agent, let it read your PR and CI history for recurring corrections, and review the improvement PRs it opens against your rules, skills, and verifiers.