The Stage: Federated Layer for Multi Agents

Mike Anderson, CTO of Covia Labs opens with a historical parallel:

Back in the dawn of the industrial era, when people first harnessed electric power, there was no electricity grid. Factory owners built their own generators. They designed their own machines and figured out how to fuel them in isolation. It took time before the grid evolved and connected everyone. We are currently in the generator phase of artificial intelligence.

Everyone is building their own self-contained agents. They run in a controlled space on their own hardware with their own data.

But to Covia, the world is in reality, federated. To do real work, agents from different companies need to work together across organisational boundaries.

"You can't simply assume that you run every single agent inside your own environment," Mike tells the room. That is why Covia is building an open-source agentic grid.

The Governance Problem

It is not enough for agents to just produce text. They have to take actions. And the problem with taking actions is that agents can do a lot of damage. Without proper controls, they can delete databases or commit major security flaws to your codebase. To him, the biggest risk with autonomous agents is not what they can do, but how we govern them.

Covia solves this through the concept of venues.

In their world, the grid is a virtual overlay network made up of these venues. Each venue can belong to a different person or organisation. By default, you have no control over anyone else's venue.

Mike explains the workflow for how agents operate safely within this grid.

1. A venue provides the agent harness to run local agents.

2. The venue assigns capabilities to agents using cryptographic tokens.

3. These tokens grant specific operations on particular resources.

If an agent tries to execute an action without the required capability token, the action fails.

In other words, an organization can own multiple venues, and in each venue, you get to define what equipments are inside, the rules behind what you can do in the room, and only people with the access cards to each venue gets to go in, much like how we will do it in the physical world.

The 25 Percent Secret

When you build complex environments, you run into the context window problem.

If you have spent time building with language models, you probably know the story. An agent starts off doing wonderful work. As you add more instructions and tools on top, it degrades. As soon as you fill up its context window, performance drops rapidly.

"The dirty secret with the model builders is that agents can only really effectively use about 25% of their context," Mike admits. This is why single, general-purpose agents fail.

The solution is specialised multi-agent teams, where each agent has very specific roles. A lower-level agent can use a cheaper model to perform a single task efficiently. Then you ask a powerful agent running a complex model to act as the manager.

But to make this work, these agents need global identifiers. Covea uses W3C DID standards to create globally addressable identities for every agent, asset, and operation. This is paired with cryptographic audit trails. Every single operation that flows through the system gets recorded and signed into a massive Merkle tree.

This then allows you to trace back exactly what happened and who authorised it.

Introspection on the Command Line

Mike switches his screen to a live console to show what happens when you combine an external model with a local grid agent. He pulls up Claude code talking to his local Covia venue. Inside the venue, he has an agent named Bob and another named Poet.

Mike types a prompt asking the poet to write a poem about the grid.

Claude does not have direct access to the API keys. It only knows the keys exist in a secret store. It reads the available tools and instructs the Poet agent to generate the verse. A few seconds later, Shakespearean verse appears on the screen.

"It is just looking at the grid and the tools available, and the context graph that the grid makes available, and it's figuring out how to do this," Mike says.

He calls this introspection. Agents can interrogate their own state and history. They can see what other agents are doing in real time. They can even clone each other.

Mike asks Claude to make a copy of Poet and name the new agent Freddy. Within seconds, Freddy is alive on the grid and introducing himself.

Pushback from the Room

The idea of a shared grid sounds great in theory. But real-world adoption is messy.

A developer in the second row who spent years building identity layers for hospitals in India raises his hand. He points out that when he tried to get different hospitals to trigger each other's agents, they hit a wall. The adoption was not a tech problem, but bureaucracy. The hospitals refused to share data because they were afraid the other hospital would poach their patients.

"How do you go on solving that particular mess?" the attendee asks.

Mike points out that this is exactly why the grid must be federated.

"As a technology provider, you can't go about telling organisations how to manage their data or their systems," Mike says. "If you try, you end up getting locked up in bureaucracy and the arguments go on forever, and this is just bad business"

Federation in this context means each organisation sets its own rules. They choose which capabilities they grant to external agents. They keep control of their infrastructure. They do not have to put all their trust into a single cloud provider.

Another attendee asks the obvious crypto question. If you are using decentralised identities and cryptographic hashes, why not just build this on a blockchain layer 2 network?

Mike and Chirdeep having built Ocean Protocol, know the crypto space well.

"Blockchains are not actually the right place to do this kind of good stuff," Mike replies. Public chains are too expensive and the performance is totally unacceptable for the hundreds of thousands of transactions an agentic grid requires.

A Synthetic Workforce

Towards the end of the session, the conversation shifts to what this actually looks like for a normal company.

An attendee asks who is supposed to manage all of this. Is it the IT department? Department heads? Managing one agent with any amount of granularity is already a headache, so managing thousands of specialised agents across multiple departments sounds like chaos.

Mike agrees we are not quite there yet. Eventually, we will need management interfaces where even non-technical department heads can view their synthetic workforce and prompt for changes.

The grid is still in early development. It is open source, and Mike invites everyone in the room to pull it down and try breaking it. The underlying technology is fast, but the real test is whether large organisations are actually ready to let their agents talk to the outside world.

---

Missed out last week? Don't worry, these conversations happen every Friday at SQ Collective.

Usually over laptops. Sometimes over pizza.

Join the next one.