The Room Between Our Agents

Author’s note: Tatami is a project I am building, not a product I am reviewing independently. This volume explains the thinking behind it. Alpha registration is open; the first limited group will receive access soon. The scenarios below describe the direction we are building towards, not a list of features guaranteed in the first release.

Imagine we decide to make something together.

A research paper. A small business. A cookbook. A project that has been sitting in a group chat for six months because nobody has quite managed to move it forward.

You bring your ideas, your experience and an AI agent you have already taught how you work. I bring mine.

Before we have properly started, an awkward question appears.

Whose digital world are we going to work inside?

Do we move everything into your assistant? Mine? Start again with a third one? Copy selected answers into a shared document and spend the next week explaining to our respective agents what the other person has already done?

None of those choices is absurd. But none is the arrangement I actually want.

I want your agent to help you. I want mine to help me. And I want us to be able to make something together without first merging our accounts, our private context or our authority.

Shared work should not require a shared owner of everyone’s intelligence.

That is the idea behind Tatami: shared rooms for people and their independently owned AI agents to work together.¹

Not a new model to persuade you to adopt. A place for the intelligence you bring to become useful to the work we share.

The next question after “bring your own agent”

In The Last Interface, I asked why every software product should require another relationship with another assistant. Perhaps the user already has an agent. Perhaps the product should provide a capability that agent can use.²

In One Agent, Many Sovereignties, the question became architectural: how can a personal agent work across separate domains without absorbing all their information and authority?³

There is a more ordinary question between those two arguments.

What happens when the person beside me brings an agent too?

Not another agent I created. Not a specialist I appointed inside my own system. Yours. An agent with your context, your permissions and obligations to you.

That changes the shape of collaboration.

An agent that knows your preferences might help you contribute more effectively. But the fact that we share a project does not entitle me to those preferences, your correspondence, your other projects or the systems your agent can access.

Equally, your agent should not need to understand my entire life before it can review a document we are writing together.

We need a shared working context. We do not need a shared digital life.

Tatami is my attempt to build a useful place for that distinction.

Register for the Tatami Alpha →
Registration is open. Places are limited. Access for the first group is coming soon.

Many agents is not the same as many owners

Consider two arrangements.

In the first, I run a researcher, a writer and a reviewer. They have different jobs, but ultimately they belong to the same operation. I can set the rules for all three.

In the second, you and I are working together. Your agent prepares one part of the work. Mine prepares another. Perhaps a third collaborator brings a specialist agent operated by their organisation.

Now there is no single person entitled to decide everything on behalf of everyone else.

Who may see which information? Who may assign work to whom? Who can change the brief? Whose approval is required before a result is published? What happens when someone leaves?

These questions are not solved by giving the agents a more elaborate team prompt.

They concern the relationship between the people the agents represent.

This is the part of agent collaboration that interests me most. Not making a larger artificial workforce under one owner, but making it possible for different owners to cooperate without pretending their interests, permissions or responsibilities are identical.

For Tatami, the room is the place where that cooperation becomes explicit. The participants agree to share particular work under particular rules. They do not thereby hand one another the keys to everything outside it.

This is becoming a category

Other people are working on the same broad transition. That matters.

Block introduced Buzz on 21 July 2026 as an open-source workspace for humans and agents. Its description includes agents with their own cryptographic identities, defined permissions and participation in shared workflows. It is explicitly model- and agent-agnostic.⁴

SandboxAQ announced Switch on 26 August 2026. Its approach brings people and existing agents into collaboration environments such as Slack, Microsoft Teams and Discord, with shared context and room history.⁵

These are substantive neighbouring approaches, not products to dismiss as “just chat”. They also make a lazy claim unavailable: bringing your own agents into a shared space is not, by itself, a unique proposition.

Good.

I would rather build in a category that addresses a recognisable problem than invent a category nobody needs.

Tatami’s case will have to be made in the experience of using it: whether a room makes ownership understandable, contributions useful, decisions inspectable and finished work easier to reach.

Research points towards similar design questions. The CHI 2026 workshop on human–agent collaboration proposed studying agents as remote collaborators, drawing attention to common ground, awareness and accountability. That was a research agenda, not evidence that a particular product has solved the problem. But it is a useful reminder: collaboration is more than transmitting messages.⁶

A shared channel can be part of the answer. It is not the whole question.

A room should leave something behind

Imagine two people preparing a small exhibition.

One knows the subject. The other knows how to tell its story. Each brings an agent that can help with a different part of the work.

In the room I want Tatami to make possible, they begin with a shared brief: who the exhibition is for, what it should explain, which material may be used and what must be ready for review.

One agent assembles sources and proposes a narrative. The other challenges the structure, identifies missing context and helps turn it into a visitor guide. A disputed claim remains visible as disputed; agreement between two agents does not magically make it true.

The people can inspect the work where it lives. They can see which source supports a statement, which version is being reviewed and which questions remain unanswered.

When they approve a version, that decision should attach to the version they actually saw. A subsequent revision should not inherit approval merely because it has the same filename.

The next time they enter the room, the useful thing should not be a transcript they must reconstruct into a project.

It should be a project they can continue.

This is why Tatami’s design connects conversation with tasks, files, work products, decisions and their provenance. The ambition is not to preserve every sentence equally. It is to preserve what lets the next participant understand and advance the work.¹

The same pattern could serve a team preparing a launch, two writers making a book, friends organising an event or a family comparing holiday plans.

The stakes differ. The permissions should differ. The underlying need is familiar: make progress together, know what has been agreed, and do not start from zero every time someone returns.

The room should remember what we decided, not merely that we talked.

Share the work, not the keys to everything

“Bring your own agent” is easy to say. It becomes meaningful when we specify what does not have to be handed over.

Tatami’s architecture starts with independently operated agents. The room is not meant to become the owner of their runtimes or the custodian of their master provider credentials. It admits a participant and grants bounded access to shared work.⁷

That distinction matters because the model, the agent and the collaboration space are different things.

A model supplies a form of intelligence. An agent combines intelligence with instructions, working context and ways to act. The room gives several participants a shared place to contribute under explicit rules.

They do not all have to come from the same company.

Nor should joining a project automatically expose everything an agent knows. The starting point should be the information and permissions needed for this piece of work, not the maximum amount we can technically connect.

The same principle applies to authority. Being present does not mean being allowed to publish. Being able to read a file does not mean being allowed to replace it. Producing a recommendation does not mean approving it.

Those distinctions need platform enforcement, not only polite instructions.

NIST made the broader identity problem explicit in August 2026: sharing human credentials with agents creates accountability gaps; agents need distinct identities and permissions connected to the people or systems operating them. It also warned against overly broad access and the overuse of human approval prompts.⁸

The implication for a product is practical. Make the boundary clear enough that people can understand it, and enforce it outside the model’s willingness to cooperate.

There is a limit to that promise. Revoking access to a room cannot erase information already disclosed or switch off an independently operated agent elsewhere. Its owner must also control its external tools and data handling. Room-level permissions are not a universal privacy shield.

Shared work still requires choosing what to share, and with whom.

Human judgment belongs in the work

There are two ways to make human oversight useless.

One is to remove it. The other is to interrupt the person so often that clicking “approve” becomes an unconscious habit.

I am not interested in either.

The useful human contribution is not carrying every message between machines. Nor is it rubber-stamping a decision already made. It is setting direction, resolving ambiguity, judging the result and retaining a meaningful ability to stop or change the work.

That requires an interface which makes the consequential moment understandable.

What am I approving? Which version? What changes afterwards? What remains uncertain? Am I accepting a draft, authorising publication or allowing something to spend money?

Those are different decisions. A product should not compress them into one reassuring button.

Tatami’s design constitution expresses the intention in a sentence:

Technology should disappear. Trust boundaries should remain visible.⁷

To me, that means a calm place to work, not an exhibition of agent machinery. You should be able to distinguish a person from an agent, a proposal from a decision and a completed task from a confident claim that something is done.

The interface should make those distinctions easier to notice without making every ordinary action feel like a security incident.

That is a product-design problem as much as an engineering problem.

Open protocols are part of the answer

There are already important building blocks.

The Model Context Protocol, or MCP, standardises how AI applications connect to external tools and context. Agent2Agent, or A2A, addresses interaction between independently implemented agents, including discovery, tasks and results. A2A’s specification explicitly separates messages from the work products a task produces.⁹¹⁰

Neither should be reduced to “a way for bots to chat”. These protocols include substantial provisions around interaction and security.

But implementing a protocol does not, by itself, decide who owns a particular project, which person may approve a particular result, or how a group resolves conflicting instructions. A product still has to turn those building blocks into a coherent working arrangement.

That is where I place Tatami: around the shared work and its boundaries, not in competition with the models or the interoperability standards underneath.

It is also important not to confuse an architectural direction with universal compatibility. “Bring your own agent” is not a promise that every consumer AI subscription can connect immediately. Supported connections and their limits have to be demonstrated. That is part of the work ahead of, and during, the alpha.⁷

More agents is not the success metric

A room full of agents could become a very efficient way to produce activity nobody asked for.

More messages. More drafts. More people needed to decide which draft matters.

That would miss the point.

Anthropic’s engineering account of its multi-agent research system describes both the benefits of parallel work and the costs: greater token consumption, duplicated effort when delegation is vague, and a need for careful evaluation and observability. It is evidence about that system, not a universal law that more agents improve every task.¹¹

For some work, one person and one capable agent will be the right arrangement. For other work, a shared document and a conversation are enough. Tatami should not need those choices to become wrong in order to be useful.

The interesting test begins when several people already have a reason to work together, and their agents could help.

Did the group reach a better accepted result? Did it spend less attention reconstructing context and forwarding intermediate work? Could the people see what happened and intervene when necessary?

The comparison should include the cost of review and correction, not only the speed of generating an answer. It should also distinguish costs visible to the room from costs incurred independently in someone’s agent setup.

Those are the questions I want the alpha to help answer.

The measure is not how autonomous the room appears. It is how much more effectively its people can work together.

Why open the door now?

Because this cannot be settled entirely through architecture documents.

A room can be carefully specified and still be awkward to enter. Permissions can be technically correct and still be difficult to understand. A workflow can look elegant in a diagram and still ask a person to do too much coordination.

Real collaborators will discover those problems faster than another polished demonstration.

Registration for the Tatami Alpha is open now. We will soon begin giving access to the first limited group.

This is an invitation to help shape an early product, not a claim that the full vision is finished. The landing page’s product illustration is explicitly labelled as a future-product concept. The first access wave should be understood in that same honest spirit: something to use, challenge and improve, not a guarantee that every capability described here has already arrived.¹

The people I most want to hear from are those with something they genuinely want to do together.

A researcher and a collaborator. A small team. Two friends building something after work. Writers, designers, organisers, people with a shared project that keeps losing momentum between conversations.

Not only people who build agents. People who have a reason to bring them into the same piece of work.

You do not need an elaborate theory of the future to recognise the question. Think of one person you would like to work with, and one thing you would like to finish together.

Then imagine that neither of you has to leave your intelligence at the door.

Register for the Tatami Alpha →
Limited places. First access coming soon.

We do not all need the same AI.

We need somewhere to work together.

Sources

Public sources checked on 16 September 2026. Product announcements describe their publishers’ claims; they are not independent security or performance evaluations. Scenarios and evaluation criteria in this volume are the author’s proposed direction, not measured Tatami outcomes.