research
From the lab
Original research and slower thoughts, written the way we'd explain them to a smart friend.
evolutionary-systems
Evolutionary & Self-Improving Systems
Code that changes itself, and search that finds answers we did not hand it.
Code That Beats Me
We built a system to search for a better way to multiply matrices, and it found one that no person on our team designed.
Read →Agents Judging Agents
In Darweel, the agents grade each other's work in a voting arena, which lets selection run round after round without a human scoring every answer.
Read →Reading Code No Human Wrote
An algorithm that emerged from search is full of steps that look arbitrary until you trace them, and trusting it means a different kind of reading.
Read →Reproduction by Merging
In Darweel, two candidate programs have children by fusing their prompts and memories, and that turns out to be a real mechanism for new capability.
Read →The Fitness Function Is the Whole Game
Whatever you choose to measure is exactly what an evolutionary system will give you, so getting the fitness function wrong means evolving the wrong thing perfectly.
Read →multi-agent-systems
Multi-Agent Systems & Orchestration
What happens when many agents work, talk, and hand off to each other.
From Writing Code to Running a Fleet
The IDE was built for one person typing into files, and that shape quietly falls apart the moment your coworkers are agents.
Read →A Backend-Agnostic Control Plane for Agents
The control plane that runs a fleet of agents should not know or care which model is behind any of them, and building it that way changes what it can survive.
Read →An Org Chart of Agents
Once agents can hire, fire, and escalate on their own, the hard problem stops being how smart they are and becomes who is allowed to do what.
Read →Demote the Model to Coach
A large language model is too slow and too expensive to drive a body in real time, so in Hivemind it stopped driving and started training the thing that does.
Read →Messaging That Never Interrupts
When one agent needs to tell another something, the naive move is to interrupt, and the interruption is exactly what wrecks the work.
Read →Opus, Sonnet, and Haiku as One Team
Not every job in a multi-agent system needs the smartest model, and building one that mixes tiers taught me where the sharp edges hide.
Read →voice-ai
Voice & Conversational AI
Latency, turn-taking, translation, and the hard parts of talking to a machine.
Three Seconds of Silence
A caller says hello and hears nothing back, and that small dead pause is where most voice AI quietly loses people.
Read →What a Call Center Learns When Every Call Is Read
For most of history a phone call vanished the moment it ended. When every call becomes text you can search and score, the whole floor changes shape.
Read →Cascade or Speech to Speech
Two ways to build a phone agent, and a plain account of when to chain three models and when to hand the whole conversation to one.
Read →Cutting the Pause Before Your Agent Speaks
A phone agent says the same lines over and over, so we stopped paying to generate them twice and the pause got shorter.
Read →Giving a Voice Agent Hands
A voice agent that can only talk is a very polite dead end, so the real work is letting it book the slot and send the message while the caller is still on the line.
Read →Guardrails Over Prompts
A business owner's first question about an AI that answers the phone is not what it can do, it is what it will never do, and that question deserves a better answer than a paragraph of instructions.
Read →One Integration, Any Carrier
A telephony provider changes one field name in a webhook and your whole voice product breaks, unless you built a seam between your code and the carrier.
Read →Running Stateful Calls on Serverless
A phone campaign is a long, stateful thing and serverless containers can vanish mid-sentence, so the whole trick is making the work survive the machine doing it.
Read →Tracing Who Is Speaking When
You cannot shorten a pause you cannot see, so we put a stopwatch on every part of a voice call and watched where the time actually went.
Read →Two People, Two Languages, One Call
Two people who share no language get on an ordinary phone call, and each hears the other in their own tongue, which turns out to be less about clever translation and more about how you split the call.
Read →agent-simulation
Agent Simulation & Evaluation
Using simulated people and traffic to test systems before the real thing.
Hiring for the Team You Already Have
If a model can read a person from an interview, you can drop a simulated candidate into a simulation of your real team and finally measure social fit.
Read →Grading the Agent Blind
An autonomous agent that grades its own homework will always pass, so I built a company with a hidden answer key and scored the agent against a truth it never sees.
Read →Simulating a Whole Company
Nobody in a company can tell you how the whole company works, so instead of interviewing everyone myself I built the company out of agents and let one agent go find out.
Read →Test Your AI With More AI
Waiting for real users to find the failure is a terrible test plan, so we built agents that call each other and break the system on purpose first.
Read →The Two Faces of Simulation
The same idea, agents standing in for something real, can stress a live system or model a room full of people, and I want to show why both are the same move.
Read →applied-ai
Applied AI in Production
War stories from putting AI to work inside real businesses, and what held up.
A Prompt and a Toolbelt Is the Product
I ran real outbound sales with a general coding agent, a big set of tools, and one instructions file, and it changed what I think a product is.
Read →Assistants That Have to Be Right
In a chatbot about movies, a wrong answer is a shrug. In one that tells a business owner how to register or file taxes, a wrong answer costs them money. That changes how you build.
Read →Compliance as an API Contract
If you want an agent to run outreach at scale, the safe move is not to trust it, but to build a gate it physically cannot get past.
Read →Make Your Product Agent-Discoverable
Agents are starting to do the searching and the buying, and a website built for human eyes is invisible to them.
Read →One Tool Per Server
Building an agent's toolbelt from small single-purpose servers instead of one big one, and why the small pieces are the ones you can actually trust.
Read →Running a Business With an Agent Fleet
The agents do the daily work of the business, and the hard part turned out to be governing them, not building them.
Read →Tools as Data
Most software teams ship a new product by writing new code. What if a new product were just a row in a database, and shipping it took minutes?
Read →Turning a Business Into Tools an Agent Can Operate
What it actually takes to expose a whole business function as typed tools an agent can call, and the things you have to harden before you dare hand over the keys.
Read →Turning a Phone Camera Into a Data Pipeline
A pharmacist photographs a crumpled supplier bill, and a vision model turns it into a live inventory. Here is what works and where it still fights back.
Read →perspectives
AI, Philosophy & the Frontier
Slower thoughts on intelligence, language, and being human while the machines learn.
A Living Extension of the Mind
A model trained on our language is not a rival species arriving to replace us, it is a living extension of the human mind, built from the residue of everything we have ever written down.
Read →Being Human When the Work Is Mostly AI
When most of what gets made is machine-made, the visible marks of human effort, care, and honesty do not lose their worth, they become the rarest thing in the room.
Read →Models That Adapt Themselves
A model that quietly retunes itself for each question it is asked, using a single small dial per skill, hints at something about how learning might actually want to work.
Read →Nearly Orthogonal
In everyday space you can only fit three lines at right angles, but in a very high-dimensional space you can fit an astonishing number that are almost perpendicular, and that is why a model can hold a whole world inside it.
Read →The Machine That Reads
A word is a tiny container we pack meaning into, and a model learns to read by watching how we pack it, one sequence at a time.
Read →What If Consciousness Is Evolved Chaos
Order keeps showing up out of noise, in brains and in evolving code, and it makes me wonder whether creativity and even consciousness are just chaos wearing a very clever mask.
Read →