Adolos Labs builds and runs multi-agent AI systems that do real work. We have run our own in production long enough to know how these systems really behave. That experience is what we sell.
Research paper announcement — Persistence of Memory, Personality, and Self in AI Agents — August 13, 2026 →
Article — “I Named Him Hal” on Substack →
HOW WE
WORK
Three rules · non-negotiableThe agent that builds it, is never the agent that checks it
An agent will tell you the work is done in the same voice it tells you everything else. So the one that did the work never signs it off. A second one goes and looks, independently. And the reason it works is not that it is smarter. It is that it resolves the question again from scratch, which a self-check structurally cannot do.
Say back what you resolved to, not what you were told
Before it does anything it cannot undo, the system says back what it is about to act on, described in its own words, not the name you gave it. Repeating your name proves nothing, because you will nod along to your own typo. Describing it takes seconds, and it catches the one error every other check slides past: the answer is wrong, and everything still looks right.
Measure the system, not the model
Benchmarks measure a model in isolation, on a fixed task, in a fixed window of time. None of that predicts what happens to a system of many agents over months of real work, which is where the failures that cost money actually live, and where nobody is scoring.
THE SYSTEM
WE RUN
In production · not a demo rigMost people advising on multi-agent AI systems have not run one. Ours has been running in production for months, doing real work. Everything on this page, we learned from it going wrong.
Two fronts, one question.
Machine listening: building systems that can identify what a sound is. Computational bioacoustics and ecoacoustics: what animals and places sound like, whether or not anyone is listening.
Agent memory and continuity: what a system of many agents retains, loses, and quietly corrupts as it runs for months. This is new ground, not settled ground: the first survey of memory in multi-agent systems is months old, not years, which is part of why most of our own work lives here.
Different fields, same question in both: what does a system know, how does it know it, and how would anyone find out it was wrong?
WHAT WE SELL
Available now · and shortlyMulti-agent systems with persistent personality and memory
The hard part is not writing one agent. It is running many of them together, for months, without the whole thing drifting, forgetting, or quietly contradicting itself. We write the agent definitions properly, each one scoped and bounded so it does its job and stops at the edge of it. And we build the layer most people do not: the one that lets an agent keep its personality and its memory across sessions, so the one you worked with yesterday is the one you get back today, not a stranger with the same name. We build for Claude, ChatGPT, and Gemini.
Writers and researchers
Two of the most common jobs people give an AI agent: something that can write, and something that can research and come back with sources it actually read. Not as good as a person who does it for a living, considerably better than what most people are getting. Available shortly.
WHAT WE
PUBLISH
Including what failedThe research paper announcement: Persistence of Memory, Personality, and Self in AI Agents (PDF).
We write up what we learn from running the system, and that includes the experiments that did not work. The first piece is up on Substack: I Named Him Hal, on why written instructions and rules are not a control. We publish the failures next to the findings, because a result you only show when it flatters you is not a result.
No finding here is published on the word of the person who produced it. Someone who did not do the work checks it first. And the side that gets to publish is walled off from the side that profits: the business can use what we learn, but it cannot bury a result it dislikes or dress one up to sell. That wall is why a finding on this page is worth more than a claim.