Agent Clinic is hands-on help for engineers, teams and founders whose agents stall, loop or say “done” when they aren't. We fix the setup in your repo, so the work finishes and you can check it.
Free · 15 minutes · Google Meet · English or German
Jackson Ly
Lead engineer at a YC-backed startup
1,029
PRs merged in 30 days
All public
GitHub and write-ups
Checks passed
run outside the agent
A lot of what goes wrong with Claude Code and Codex comes from the setup around the model: context that grows until the agent forgets, checks the agent writes for itself, and runs without a stop condition. That part is fixable.
97.9%
of the tokens in my busiest month were the model re-reading its context. Context size is a cost lever.
16.7%
of my follow-ups to agents came from the agent falling short. The other five in six were not agent failures.
17 of 26
statements in one post needed correcting by a fresh reviewer, after my writer agents had reported green.
Numbers from my own runs. Read the full write-up
BeforeAfter Agent ClinicFree · 15 minutes · Google Meet · English or German
Pick the symptom that sounds most like your week. This is the first thing I'd check, before we've even talked.
It says done, but it isn't
Likely causeThe agent wrote the check it then passed, or nobody defined "done" before the work started.
First thing I'd checkAgree on the check before the task, run it outside the agent's session (a hook or CI), and let a fresh reviewer open the sources.
In a session we look at your actual runs and fix it in your repo.
Book a free callEvery session works on your repo and your runs. These are the areas that come up most.
Module 01
Instruction files short enough to be followed, permissions that stop asking about safe commands, and hooks that run your checks without you.
Typical results
Module 02
Checks the agent didn't write itself, and a fresh reviewer that opens the sources. You read results instead of every diff.
Typical results
Module 03
Briefs sized for one run, and stop conditions that end long runs before they drift for hours.
Typical results
Module 04
Worktrees, one task list and state kept in files, so a usage limit or a crash loses nothing.
Typical results
Module 05
Where your tokens go (context size, cache, which model does what), and what can safely run cheaper.
Typical results
Module 06
One setup in the repo and review rules for agent pull requests, so results stop depending on who ran the agent.
Typical results
Not sure which of these is yours?
That's what the free call is for.
Agent Clinic is new, so there are no client logos here yet. These are my own runs, each with the fix I'd bring to yours.
Free · 15 minutes · Google Meet · English or German
Courses, prompt packs and AI gurus
Agent Clinic
Free · 15 minutes · Google Meet · English or German
I did a business degree first and became an engineer by building. Today I'm lead engineer at a YC-backed startup. On my own products, most of the code is written by coding agents I direct, often several at once and overnight.
In my busiest 30 days, my agents and I merged 1,029 pull requests across 21 repositories. I wrote up what that taught me, including what went wrong, and I build Morndo, an open-source tool for running many agents.
Agent Clinic is for the problems I had to solve myself, usually after a run that went nowhere.
Step 01
Pick a time on cal.com. Google Meet, English or German. Bring the run that went nowhere if you have one.
Step 02
If I can help, you get a short proposal with a clear price. If I can't, I'll say so and point you somewhere better.
Step 03
We work in your setup, with the tools you already pay for. You keep everything: the config, the checks and the notes.
Free · 15 minutes · Google Meet · English or German
No. Nobody can honestly promise that. What I can promise: after 15 minutes you know whether I can help, and you get a clear price before any paid work starts.
You can, and plenty of people do. It took me a lot of wasted runs. If fifteen minutes saves you some of that, good. And nobody here gets told "skill issue".
A course teaches the features. We work on your repo and your runs, and you leave with a setup that is already in place.
Both. I use them every day and pick per task. I won't try to move you to a different tool.
No. We can work on a screen share in your environment, so your code stays with you. If you want me working in the repo, we agree on access first.
The call is free. After it, you get a clear price before any paid work starts.
Yes. The work is the same, plus a setup everyone shares and review rules for agent pull requests.
The tools change every month. Briefs, checks, stop conditions and state kept in files have carried over every update I've been through.
Fifteen minutes is enough to see what's going wrong and whether I can help. If I can't, you still leave with a next step.
Free · 15 minutes · Google Meet · English or German