
How we build eLLMo AI with people and agents as one team
Your people and your agents share one live view of the work, a person steps in at the moment it matters, and nothing stops when a laptop closes. We built that place for our own work on eLLMo AI. It is called Kojo, and your team can build in it too.
Your people and your agents can build as one team, with one live view of the work for everyone. Your agent is partway through a change when it reaches a step that matters, and it asks before going on. The teammate who knows the answer sees the question on their own screen and approves it, and the work moves on instead of waiting for a review at the end. At the end of the day, the person who started the task closes their laptop, and the agent keeps working.
That is how we build eLLMo AI today. Until October, each of us built it beside our own agent.
Why we needed it
An agent would stop to ask a question that only a teammate could answer, and the work sat still until the two of them found each other. The person beside the agent was often not the one who knew.
The more agents we ran, the more changes came back to us at once. Adding an agent takes a minute. How much a person can read, weigh and decide in a day stays the same. So the work piled up in review, waiting on the one thing only people bring: judgment.
Long tasks stopped when the person who started them went home, because the agent lived on a laptop that went home too. A cloud machine kept on overnight cost as much as one working at noon, and one we switched off had to set everything up again the next morning.
We ran into all of this every day while building eLLMo AI, so we built one place that solves all of it at once. We named it Kojo, after 工場 (kōjō), the Japanese word for factory. Kojo is an AI-native factory: a swarm of agents and people building software together in one shared place.
What changed for us
We built Kojo in the middle of our own work on eLLMo AI, and each stage after the idea made that work faster. Now we build eLLMo AI in it every day. At this pace, we ship more than 5,000 pull requests and run more than 140 billion tokens a year.
From November to April, we did AI engineering and shaped the idea for Kojo. From May to September, we engineered a software factory on self-verifying goals, loops and our own skills. A typical week brought 92% more contributions than in the first phase. In October, Kojo arrived, and our engineering became swarm-native: people and agents working together in one live session. A typical week now brings 796% more contributions, nine times our pace in the first phase.
This is what a team of four people can do with the right vision, the right strategy and the right tools.
The bigger change shows in our days. A question reaches the right teammate while the work is still moving, and a task started in the afternoon keeps going after we log off.
How we think about building with agents
Kojo is an ADE, an agentic development environment: the place where people and agents build software together.
We think about it the way we think about our own product. eLLMo AI sends a panel of simulated buyers through a store, each with its own habits, and together they show where the store loses people and why. We build software the same way, with many agents on a problem at once, each working from its own angle.
A swarm like that needs people on the floor. Ours stand in the middle of the work, where they set the direction and answer the agents' questions as they come. The agents explore, and our people decide what ships.
One session runs one agent, so a problem with several angles gets several sessions side by side, and our people move between them.
What we built
Working this way took four pieces of infrastructure, and we built all four.
- A computer for every session that sleeps and remembers. Whenever nobody needs it, it sleeps, and it wakes with your server still running and your terminal where you left it. If the computer is ever replaced, the next run starts a new one from the pushed branch and the saved conversation.
- The whole team, live, in one session. Everyone shares the conversation, the computer and its terminals from a browser on a desktop or a phone, and each person's cursor shows where they are working.
- An agent that works while nobody watches. It pushes its branch as it goes and asks before the steps that matter, and a question nobody answers in time ends in a no, never a quiet yes.
- A person who decides what merges. Only a person chooses the repositories a session works in. Changes come back as draft pull requests in your own GitHub, and nothing merges until a person approves it.
Together, they turned a team of people, each beside one agent, into one team building with many. They carry our own work every day, and now they can carry yours.
To build this way with your own team, visit kojo.build or start a session in Kojo.
See this in action on your page
eLLMo sends simulated buyers through your product page and returns a ranked list of what stops people from buying, in their own words.
More from the blog