Topos: the place where agents work in plain sight

May 16, 2026

We are announcing Latere Topos. Topos (τόπος) is Greek for place: where work happens, in view. The job is narrow on purpose:

A platform where every AI agent has a name, a clear set of permissions, and a full record of what it did. Set up once, watched live, reviewable after, and reused by every product that builds on it.

The trade nobody agreed to

AI agents take work off your hands, and often out of sight. The cost, as things ship today, is that the work disappears into a place you cannot see.

You get an answer and a bill. The steps in between stay hidden. One agent quietly started six more, and you have no way to see them. One of them used a tool it was never supposed to touch, and nothing recorded the attempt. A draft went to an executive with two customer records still in it, and nothing flagged the leak. The rules meant to prevent all of this are buried deep in the code, readable only by the handful of people who wrote them.

This is the quiet trade most agent setups make: more autonomy, less visibility. Topos refuses it. Our bet is that an agent you can watch beats an agent that is slightly faster, and that the gap only grows the moment you run more than one.

An agent is a document

The first decision is what an agent is. In Topos it is a document: a name, a human owner, what it is allowed to do, the model it uses, the tools it can reach, a budget, and the approval steps that apply to it. Everything past that is just the work it does day to day.

That sounds like paperwork until you notice what it buys. You can read the document before the agent runs and again after it finishes. You can compare versions side by side. Every change is signed, so you always know which person made it. You can hand it to someone in compliance who has never seen your code, and they still get the whole picture. The agent is something you can hold in one hand.

Visible work

Every action is a message, and the message is the work. The system records each one as it happens. Using a tool, starting another agent, removing sensitive details, refusing a request, sending something out: each one is written into a single running log for that session, with a timestamp, who did it, and whatever it produced attached right where it happened. You can watch it live, or read it three weeks later and rebuild exactly what happened and who caused it.

One session can branch out across a dozen agents and three layers of helpers. A simple map shows the whole family tree at a glance: who handed work to whom, who is still running, who stopped and exactly which rule stopped them. Every step is on the record. If it happened, it is in the log; if it is not in the log, it did not happen.

Power is declared up front

What an agent is allowed to do is spelled out in plain text. Internet access, files, connected tools, the right to start other agents, spending limits: each one is listed separately, kept under version history, easy to compare, and reviewed by a person before anything runs.

Permissions can only shrink as work is handed down, never grow. When one agent starts another, the new one gets the same permissions minus whatever was taken away, and it can never grant itself back what it lost. Start a helper with internet writing turned off (-network:write) and that ability is gone for everything below it, for good. A request for something off the list comes back as refused (denied) and stops right there. Every refusal shows up as its own line in the record, right next to the exact rule that caused it.

The handoff stays human

Letting an agent run on its own is right for some work and reckless for the rest, and that line keeps moving. Topos lets you place a person exactly where a decision matters: an approval step before something is sent, a review to remove sensitive details before an executive reads it, a sign-off before anything reaches the open internet. The agent pauses until that sign-off arrives. Full freedom when it is earned, full control when it is not, the same way of working in both cases.

AI executes. Humans decide. The product exists to keep that sentence true under load.

Review before you see it

Some checks do not need a person. Before an agent's work reaches you, Topos can put it through an adversarial review: an independent critic cross-examines the output, the agent has to fix or defend every objection, and only the disputes that survive are handed up. It sits one step ahead of the human handoff, so the work a person signs off on has already been argued against. Adversarial Review runs this against a coding session or a diff, vendor-neutral and auditable end to end.

A substrate other products consume

Topos is where agents are set up. The same group of agents, carrying the same names, permissions, and record of what they did, can then be used across a team chat, a coding assistant, an internal app, a support-ticket tool, a workflow system. Set an agent up once and every place that picks it up inherits the same rules and oversight, so each product builds on one strong copy.

It stands on the rest of Latere: Cella for the safe space an agent runs in, Lux for model access that keeps provider keys away from every process, and Latere Identity so an agent's owner is a real account with a real name. Topos is the layer that turns those building blocks into something a team can run with confidence.

Try it

The console is in private preview at topos.latere.ai. Identity, permissions, the live conversation, the map of who is doing what, and the full record are the screens you work with; an agent is the document that ties them together.

If you are running more than one AI agent and have started to feel the blind spots, the agents you cannot see, the refused requests nobody wrote down, the permissions only a few people can read, reach out. We are especially interested in teams already running a group of agents who have an incident story to tell.

Topos is a product of Latere. Founded by Dr. Changkun Ou.