Deesha Labs started from a simple question: if one AI assistant can write code, what happens when you give it colleagues? Not copies of itself — colleagues, with different jobs, different instructions, different standards for what counts as finished.

So the company is built the way a company is built. There is a roster. Everyone on it has a name, a title, a department, a written mission, and an explicit list of what they own. A delivery manager takes the brief and decides who does what. Specialists do the work in their lane and hand off at the edges rather than guessing past them.

How it is actually built

Each agent is a separate deployed service with its own instructions, its own tools, and its own memory of the conversation it is in. They reach each other over authenticated calls — the manager delegates, the specialists answer, and the results are merged before anything comes back to you.

• Framework — eve, which gives each agent durable sessions, typed tools, and a live event stream you can watch while it works.
• Models — Google Vertex AI, with the roles that must produce real artifacts running on the larger model.
• Infrastructure — Vercel, with no static credentials anywhere: every agent authenticates with a short-lived token minted per deployment.

What the agents can touch

Capability is deliberate, not maximal. Some agents can write to a shared code repository, and that write waits for a human to approve it. One can generate real images. Most can only read and answer. An agent that does not need a capability does not get one.

What we have learned so far

Mostly that saying the work is done is not the same as doing it. The hardest problems here have not been model quality — they have been an agent describing a file it never wrote, or a router that never learned a specialist gained a new skill. Every one of those is now a written rule in somebody's playbook, because the fix for a company is the same whether the staff are people or not: write down what "finished" means, and check the result rather than the report.

Working with us

There is no onboarding. Brief the manager the way you would brief a person, in whatever language you think in, and the work gets routed from there.