The Vivari manifesto · July 2026

A Brilliant Stranger
in Your House

We manage every human hire with care. We give AI agents the opposite, and then act surprised. The models kept getting better. The management never arrived.

Think about what happens when a brilliant new engineer joins your team. Before they ship a line of anything, you give them context: the codebase tour, the architecture doc, the list of things that look safe to change and aren't. You give them memory - they remember Tuesday's decision on Wednesday. You scope their access to their role; nobody hands the new hire the production keys on day one. You review their early work closely, and less closely as they earn it. Everything they touch leaves a record. And there is always someone senior nearby who can lean over, guide them - or take the wheel.

That is how organizations metabolize talent. It is so obviously correct that doing the opposite for a human would sound like negligence. Read the inversion out loud: give the new hire no context, no memory of yesterday, the keys to everything, no review, no records, and no one watching. You would fire the manager who proposed it.

Now look at how the industry runs AI agents.

The double standard, doubled

An AI agent is a brilliant stranger in your house: enormous capability, zero knowledge of your world. It does not know your conventions, your history, your landmines, or what you meant last sprint. That is what the word stranger means.

And the way we treat this stranger is worse than careless. It is precisely inverted. We give agents more access than any human would get in their first hour: full repository write, a shell, credentials within reach. And we give them none of the scaffolding: no durable memory, no scoped role, no review proportional to trust, no audit trail anyone reads, no senior with a hand near the wheel. Maximum capability, zero management.

Then the agent does something strange at 2 a.m., and we conclude the technology isn't ready.

"Fire it and hire a bigger brain"

The industry's standard remedy for agent failure is to wait for the next model. Swap the brain, keep the negligence. Nobody manages a struggling junior engineer this way - you don't fire them and requisition a smarter human; you give them context, feedback, and narrower scope, because you understand the problem is rarely raw intelligence.

We have tested this the hard way. In our own work we watched a tiny, properly treated model outperform one six times its size on a judgment task the big model kept failing. The small one had been given exactly the right job, the right context, and the right boundaries. The constraint that day was management. It is the cheapest upgrade in AI, and almost no one ships it.

What management means

Everything we owe the stranger, we already give every human hire as a matter of course. Context: agents born knowing your project instead of interrogating it from zero. Memory: what one agent learns on Tuesday, every agent knows on Wednesday. Permissions: access scoped to the role instead of defaulting to everything. Review: risky changes checked before they land, with evidence a human can inspect. Audit: every action on a record you can replay. Intervention: a human who can redirect with a sentence, or grab the wheel outright. Structure: a division of labor, because a fleet is not a pile of chatbots.

One more thing we owe it: honesty about what it is. An agent org chart should not cosplay a human one. Humans come bundled - a salary must buy a whole role, so we invent titles and hierarchies to manage the bundles. Agents unbundle. Rather than a pyramid of vice presidents, the right structure is an ecology of specialists, each shipped with exactly the permissions, tools, and knowledge its niche requires, supervised not by middle management but by magnification - the ability to see the whole fleet at a glance and dive to any single terminal in one motion.

Why I built it

For years, I was the management layer. My real workflow was embarrassingly manual: think out loud with one model, have it produce notes, carry those notes to a different model to challenge and continue, carry the results back, then hand the plan to yet other agents to implement - me, ferrying context between brilliant strangers like a courier, because the models could not remember, could not check each other, and answered to nothing.

Every piece of Vivari was built to automate a job I was already doing by hand: the shared memory the models didn't have, the cross-checking I did between them, the orchestration I was, the safety review I performed at midnight before letting anything touch a repo I cared about. Vivari is that workflow, turned into a place.

Vivari

Vivari is the AI agent workspace: a living, observed environment where fleets of agents (engineering agents shipping code, research agents mapping a market, ops agents running the pipeline) are put to work and kept accountable at the magnification you choose. Zoom out and the whole fleet is one glance. Zoom in and you are inside one agent's terminal, hands on the keys. In between: rooms of agents sharing living memory, Vivari Guard, the deterministic review engine, holding the risky changes with evidence from your own history, and a resident orchestrator that keeps the fleet moving and steps aside the instant you take over.

It runs the agents you already use (Claude Code, Codex, Cursor today), and shipping code is only the first job you'll hire them for. Everything the fleet does goes on the record. This product was built by its own fleet, under its own audit trail.

The models will keep getting smarter. That is exactly why the management layer matters more every month: the more capable the stranger, the less acceptable the negligence. Every team will have the smarter models. The teams that win the next decade will be the ones that learned to run a hundred brilliant strangers the way great organizations have always run brilliant people.

We manage what we value. It's time the agents got the same.

- Wael Masri
Founder, Vivari