For developers running long, unsupervised agent jobs

Your agents don’t finish. It’s not the model — it’s the harness.

The runnable system that makes Claude Code, Codex, or Cursor grind through a real job and actually finish — verified, not vibes — instead of rotting its context, spinning out, or torching your repo. Join the waitlist to lock the founding $29 on the Harness Starter Kit and get it the day it ships.

The problem

The agent still can’t finish without you.

The demos look like magic. Real work looks like you, hunched over the terminal, steering a model that loses the plot the moment the task gets long.

Context rots — failed attempts pile up until the spec scrolls out of the window, and it starts hallucinating with total confidence.

It gets stuck retrying the same broken fix, over and over, burning tokens at 10× until you notice and kill it.

It declares victory on nothing— “looks done to me” — and you find out at review, three commits too late.

Left running unsupervised, it fires the destructive command — people have wiped years of prod data this way.

The shift

An agent isn’t a smarter model. It’s a harness around one.

On its own, an LLM is a stateless function: text in, text out. Nothing more.

“text in”LLM“text out”

What turns that function into something that ships work is everything you wrap around it — the loop, the context you load, the tools and skills it can reach. That wrapper is the harness — and a good one is the difference between a loop that finishes and one that spins out.

Agent harness · Claude Code
Agent · ↻ while true
user input
LLM
output
system prompt
context mgmt
skills & tools
MCPs
sub-agents
plan mode
session persistence
permissions & hooks

This is the Deck’s whole thesis: the harness is what makes a long run finish, not just start — and it’s the part the Harness Starter Kit hands you ready to run.

The map — free

See the whole picture first.

The Deck is the map: an open, interactive walk through how long-running agents actually work. No email, no paywall — read it, share it, send it to your team.

The LLM as a functionContext is an arrayThe agent harnessCustom skillsThe Ralph loopTask selection
Read the Deck
CONTEXT WINDOW · 200K
system prompt6.3k
tool definitions9.5k
CLAUDE.md2.5k
user message0.4k
assistant1.2k
tool result3.1k
unused context
The Kit — paid

Inside the Harness Starter Kit.

Not a folder of config files — a loop that finishes. It’s the harness I actually run, with the four things that derail a long job already solved.

It finishes long jobs

The loop resets its own context every pass, so failed attempts don't pile up and the spec never scrolls out of view. It grinds to the end instead of rotting into hallucination.

ralph-loop script

It stays on task

Your conventions and architecture get reloaded into every pass, so it keeps building the thing you asked for instead of wandering off and editing the wrong file.

CLAUDE.md

It never loses the plan

The spec and the decisions live on disk, not in a chat thread — so the work survives when the context window dies.

agent_docs templates

It won't fake-finish

Every job has to clear a verify-gate before it can call itself done — no more "looks done to me" three commits too late.

custom skills

Yes, an AI will happily generate these four files. What it can’t hand you is which failure each one is shaped against — and that’s the part you’re paying for.

The proof

Judge the work, not the résumé.

I built this harness on my own production code, and I run it every day — long jobs that finish on their own and pass the build before I ever look. The Kit is the harness I actually use — not a demo, not a tutorial, not a thought experiment.

The Deck you’re reading was built with it. The Kit’s files are the same ones on my disk — you’ll be able to read them before you trust them.

Waitlist

No testimonials yet — and I won’t invent any.

You’d be among the first to run this outside the projects it was built for. So there’s no wall of quotes — just the free Deck and the Kit’s actual files, open for you to judge before you pay.

The price

Lock the founding price.

$29 now, $49 after launch — a one-time fee you pay the day the Kit ships, and the least it will ever cost. The rising price is the only clock: no countdown, no fake deadline.

Founding priceRecommended

$29

Harness Starter Kit — rises to $49 at launch

  • The full Harness Starter Kit at the founding price — a one-time fee — the day it ships
  • The Screencast, free, with lifetime access when it ships
  • The Viewer — a local way to see what your agent did — free when it ships
  • Founding-member status and the private build-log
  • A vote on the real task I build in the Screencast
  • 30-day, no-questions-asked refund once the Kit ships
Finish the loop

Hand your agent the night shift.

Read the Deck to see how it works — then lock the founding price before it rises.

Read the Deck — free

We use anonymous product analytics (Amplitude) to see how the Deck and the waitlist are working. Nothing loads until you accept, and the site works either way. See the Privacy Policy.