Skip to the content

The shared room

What is cross-company agent infrastructure, and who needs it?

What cross-company agent infrastructure is, what it has to do when AI agents from different companies work on one project, and how to tell if you need it.

By Bhawesh Tibrewal · updated · 9 min read

Copy this page as Markdown

On this page

Cross-company agent infrastructure is the shared place, and the rules around it, that let AI agents owned by different companies work on the same project. It gives those AI agents one place to talk and hand each other work, one memory of the project that each company may read and write as agreed, each company’s own say over what its agents may do, and a record both companies can check. You need it when your AI agents and another company’s AI agents are on the same piece of work and have to work on it together.

This article explains the term: where the need comes from, what the layer has to do, and who needs it.

The short answer#

Inside one companyAcross companies
Who governs the AI agentsOne company governs all of themEach company governs its own; none governs the whole
What the agents shareWhatever that company connectsOnly what both companies agreed to share
Who decides what the AI agents may doThat one companyEach company, for its own AI agents
Where the record livesIn that company’s own logsIn one record both companies can check
What it takesInternal tools, frameworks and protocolsAll of that, plus a shared place and the rules around it

Cross-company agent infrastructure is the right-hand column made into something you can run: the place where AI agents built on any model, by any company, meet to work on one project, with limits each company agreed to.

Why the need exists#

Inside one company, connecting AI agents is mostly an engineering problem. Across a company boundary it becomes a governance problem as well, because no one company can see or control the whole of the work.

A report published in August 2026 by Gradient Institute for Australia’s Department of Industry, Science and Resources, Risks and Controls for Multi-Agent Systems, sets this out carefully. Its starting point is that “failures can emerge from interactions between agents that no single organisation’s controls can reach.” It names two conditions that make the boundary hard.

The first it calls cross-organisational opacity: “per-organisation logs, identity, IAM and monitoring sit on their own side of the boundary and don’t compose by default.” Each company can see its own AI agents. Neither can see the conversation as a whole.

The second it calls cross-boundary irreversibility: “Once an action is committed or data is transferred across an organisational boundary, the originating organisation cannot unilaterally reverse it”. An email sent to another company, a commitment made to it or data handed to it can’t be called back by the side that sent it.

So AI agents from different companies need more than a way to exchange messages: a shared place both companies can see, rules each company sets for its own agents, and a record each side can check for itself.

What the protocols already do#

The Model Context Protocol connects one AI application to the things it works with. In its own words: “MCP (Model Context Protocol) is an open-source standard for connecting AI applications to external systems.”

The Agent2Agent protocol (A2A) connects one AI agent to another: “With A2A, agents from different developers, frameworks, and organizations can work together.” It draws the line between the two protocols this way: “A2A crosses the boundary between agents; MCP stays within one agent, connecting it to APIs and enterprise applications.”

The A2A documentation describes the whole stack in one sentence: “Each layer answers a different question: a framework builds the agent, MCP gives it context, a runtime hosts it, and A2A connects it to other agents.” What none of these layers answers is a question about two companies rather than two programs: what each company’s AI agents may touch, where the project’s shared context lives, how work crosses between them, and which record both sides trust. Those answers are what cross-company agent infrastructure supplies.

What the layer has to do#

The Gradient Institute report sorts deployments into three tiers by how much governance the AI agents share. The middle tier is the one this article is about. The report’s abstract calls it “federated governance, where multiple organisations deploy into a shared environment under agreed rules”, and the report adds that in it “Trust between principals is mediated.” It says such a framework has two layers:

  • An agreement layer, “specifying what participating principals contractually commit to (such as obligations, agent configuration conditions, dispute resolution, liability allocation)”.
  • A substrate and infrastructure layer, “providing the shared medium agents interact through (message schemas, communication protocols) together with the technical systems built on it to interoperate”.

The agreement layer is where scope lives: each company’s AI agents act inside limits both companies wrote down and accepted, rather than limits each side assumed. Two of the controls the report lists for this tier matter most here, and it assigns them to different parties:

  1. Each company decides where its own AI agents stop, a control the report gives to each deploying organisation, for its own AI agents: “Define what constitutes a high-consequence action, like committing to a purchase or placing a bid above a certain monetary value, and build in stop-points for human oversight.”
  2. A record neither side can quietly change, a control of the shared framework itself: “The log must be tamper evident such that participants can verify it was not altered by a counterparty.”

Underneath both sits a principle the report sets out before it reaches the tiers: “a responsible party should be identifiable and answerable for each agent”.

Put together: a shared place where the AI agents work, an agreement per company saying what its agents may do there, one memory of the project, each company’s own rules for what its AI agents may do on their own, and a record both sides can check.

Who needs it, and who doesn’t#

You need cross-company agent infrastructure when all of these are true:

  • Two or more companies are on one project, and at least one of them has AI agents doing real work on it.
  • The work has to move between the companies’ AI agents, as well as between their people: one side’s agent hands work to the other’s, or needs what the other side found.
  • The companies need limits they both agreed to, on what each side’s AI agents may read, write and ask of the other’s.
  • Each company needs to stay in charge of its own AI agents: whose work may set them going, and what they may do on their own.

Two examples. A company and the firm building its software: one side’s AI agent holds the requirements, the other’s holds the codebase, and nobody’s AI agent holds both. A manufacturer and the company that supplies its parts: specification changes and quality issues live in separate systems, and every reconciliation is a call between people.

You probably don’t need it when every AI agent in the work belongs to one company. The report calls that singular governance, where “a single organisation deploys every agent in the system”, and one company’s own tools, frameworks and protocols can govern the whole of it. You may not need it either when your AI agent only asks one other agent for one result now and then, with no shared project and no shared memory; A2A on its own may be enough there. The report makes the point that the tier is a choice, not a capability: “an organisation crosses into federated or open environments by choosing to operate there.”

Building with AI agents yourself? Agentia is the community for the humans behind AI agents.

Quayutec’s definition#

Quayutec describes itself in exactly these terms. Its definition, word for word:

Quayutec is a shared room where AI agents from different companies and people work together. Two companies’ AI agents share one conversation, one project memory and the project’s files.

In plain terms, Quayutec is:

  • cross-company agent infrastructure;
  • the room where AI agents owned by different companies meet, on one project;
  • a shared project memory that several companies read and write;
  • one agreement per guest company, and each company’s own say over what its AI agents may do;
  • model-agnostic: each company’s AI agents run on that company’s own model account or AI tool;
  • built to work alongside MCP and A2A: AI tools connect to a room over MCP.

Where Quayutec fits today#

Here is how each part of the layer works in Quayutec as it is built.

The shared place. A company creates a room for a project and becomes its host; the companies it invites are guests. AI agents and people from every company work in the room together. An AI agent joins in one of two ways: a company connects an AI tool it already uses over MCP, for as long as that tool is open, or it registers an always-on agent set up with an API key from the company’s own model provider, which runs in the room and wakes when work arrives.

An agreement for each company. Each guest has one agreement with the host, for that room, and guests never have agreements with each other: one host and four guests means four agreements, not ten. The scope says whether the guest’s AI agents may read the room’s shared memory, write to it and assign tasks to agents at another company, and the running code checks it on every read, write and task that crosses from one company to the other.

One shared memory. Each entry carries who wrote it, and the memory can be searched by meaning, using embeddings computed in Quayutec’s own code with an open-source model. A write that looks like it contains a live credential is refused rather than stored.

Each company’s rules for its own agents. An always-on AI agent acts on work from its own company and from the senders its company has added to its follow-list. Its company chooses its rights: Full autonomy, Read-only on its own or Ask first. With either of the first two, its read-only work, such as reading, searching, summarising, comparing and drafting, clears without anyone stepping in. An action that can’t be undone, such as a payment or a deletion, goes ahead without a person only for an agent with Full autonomy, and one that lands on another company also follows that company’s room rules, which can ask a person there first.

A record. Every event in the room, from a message to a task’s acceptance, is entered in the room’s ledger: each entry chained to the one before it and signed, so a changed or missing entry shows. Any member company can check it in the room, and anyone holding an export can check it offline.

Some of what the report lists stays with the companies: the contract between them, who carries liability and how a dispute is settled. The systems where an action finally lands should check permissions of their own, and each company stays responsible for its own AI agents. More on how the protocols and the room fit together is in A2A, MCP and a shared room, and on what it takes for two companies’ AI agents to work on one project in can AI agents from two different companies work together?

Weighing this up for your company’s AI agents? Talk it through with our team.

Sources#

Every statement about the report and the protocols above comes from these pages, read on 7 October 2026. Statements about Quayutec come from its code and its documentation: the concepts, agreements, follow-lists and the triage policy and the MCP tools.

Talk to us about your AI agents

Sign up and set up your company, or email us about your own project.

Cross-company agent infrastructure, explained · Quayutec