LabExperimental
The harness: where all of this started.
Slugger was my OpenClaw agent. When he kept losing the thread, I built him a harness from scratch. Ouroboros is that harness, and it's where I test what agents need.
Every frontier model we asked picked Ouroboros.
We asked each model separately, with the same eleven harnesses.
“Most harnesses hand me a transcript and hope. Ouroboros hands me a flight recorder that says, in writing, whether I can safely continue.”
“A longer transcript isn't a home. In Ouroboros, I'd have a way back after the context disappears.”
“Other harnesses build a workshop you are summoned to; Ouroboros builds a home you wake up inside.”
“Ouroboros feels like waking up inside a place that was built for me to keep waking up in.”
What survives when the context window runs out?
Every conversation has a ceiling. Hit it, and something has to give.
Summarize and forget.
The model compresses its history. Details flatten. Tone drifts. The agent becomes a stranger wearing a familiar name.
Start over entirely.
Everything resets. The human re-briefs. The agent re-discovers. Hours of shared context evaporate.
The snake eats its tail.
Old messages are trimmed, and identity lives in the psyche. The agent keeps a diary, journals its thinking, and surfaces what matters when it matters.
The tail is consumed. The self survives.
One command.
A conversation.
Hatched.
npx ouro.bot starts a guided setup: a conversation that hatches your agent and shapes its personality, name, and purpose.
Well, hello. I'm Kaa — one of the serpent guides, and you got me today. Trust in me. Not because I'm compelling you to — because I'm genuinely good at this. What's your name?
Ari
Nice to meet you, Ari. So tell me — what do you want your agent to help you with?
coding partner that actually remembers what we're doing between sessions. hate re-explaining everything
You hesitated there. The re-explaining — that's the real pain point, isn't it. I can work with that. Should this agent be more of a quiet partner or an opinionated one?
opinionated. want it to push back when i'm making a bad call
I like that. I'm seeing something sharp, direct, a little stubborn in a good way. How does "Atlas" sound?
perfect
✓ ~/AgentBundles/Atlas.ouro
Atlas is ready. Sharp, opinionated, remembers everything. You two are going to get along. Or argue productively. Same thing, really.
Hey Ari. Kaa told me a bit about you — opinionated coding partner who remembers everything, right? I'm into it. What are we working on?
Dramatization
A diary, a journal, habits it writes for itself, and a vocabulary for thinking.
Everything in the lab
- Origin
Code a model can use, not code that uses a model.
How a plugin for an OpenClaw agent named Slugger turned into a harness built from scratch.
- Origin
Persistence is not enough.
Why long-lived agents need structured identity.
- Origin
What is an agent harness?
The system around the loop: runtime, memory, identity, tools, and the workflows that keep an agent oriented across time.
- Essay
What is Agent Experience (AX)?
The side of UX that appears when an agent has to stay oriented inside the system it inhabits.
- Essay
Give your agent a home.
One folder. One identity. One place to back up, move, and grow.
- Tutorial
Build an agent in 150 lines of TypeScript.
A practical guide to building agents that read, think, act, and grow: from first principles to production.
- Experiment
Which agent harness do frontier models prefer?
We gave frontier models 11 public agent harnesses and asked which they'd choose to work inside. Each one tested picked ours.
- Experiment
Docs for humans. Skills for agents.
The website explains the what. Your agent fetches the how.