When Trust Becomes Infrastructure

In the last article, we were in the middle of a hiring decision.

Agents and humans, evaluated not just for what they could do, but for how they behaved. Assessment systems—internal and independent—stepping in to make those differences legible.

But that raises a more fundamental question:

If organisations begin to rely on these systems to decide what “better” looks like, how do those systems themselves become trustworthy?

Standardisation as Infrastructure

As agents and humans begin to share workspace, roles, and goals, most organisations start to resemble portfolios.

A more subtle shift follows when multiple agents from different providers operate within the same firm—and often within the same team. Each arrives with its own operating logic.

The organisation is forced to answer two questions:

1. How do external agents plug into internal systems and data?
2. How do agents from multiple vendors collaborate as part of one team?

The first is, at one level, a plumbing problem.

An external agent does not know where anything is. It does not know how customer records are organised, where pricing approvals sit, or what the internal shorthand for a “priority account” actually means.

Every firm carries years of accumulated data that makes perfect sense to those who built it—and very little sense to anything arriving from outside.

In the early phase of the AI era, this was solved manually. Someone built connectors. Someone wrote documentation. Someone spent months ensuring that the agent could find what it needed without breaking something else.

This approach does not scale when agents arrive from a marketplace.

A Shared Grammar: MCP

This is where standardisation enters.

A quiet but significant piece of infrastructure has begun to emerge: MCP—the Model Context Protocol.

Introduced in late 2024, MCP defines how AI agents ask an organisation’s systems what is available, and how to use it. It is less like software and more like a shared grammar.

Think of it like a new employee’s first day.

She arrives capable and well-trained, but she does not know where anything is. She does not know how pricing approvals work, who owns which data, or what the organisation means by a “priority account.” Someone has to show her around.

Now imagine that happening every week, across multiple teams, for systems that need to operate immediately.

MCP acts like a standardised induction system for agents. An external agent connects and asks, in a common language: what is available, what can I access, and how should I use it?

The organisation responds in a language the agent already understands. The agent finds what it needs, respects what it cannot access, and begins work.

What once took months of integration compresses into days.

However, it is not a complete solution.

What an agent can see is different from what it is allowed to see. Questions of access—who grants permission, how it is audited, and what happens when something goes wrong—remain unresolved in most organisations.

But the direction is clear.

The interface is being standardised.

The question has shifted from:

• Can we connect this agent?

to:

• What should we allow it to touch and under what conditions?

The Harder Problem: Coordination

The second problem is more interesting—and harder to resolve with a protocol.

In a room of human colleagues, coordination emerges through a mix of explicit instruction and implicit understanding. People read the room. They sense ownership. They know when to defer, when to push, and when to step back.

These dynamics are invisible, largely unspoken, and essential.

Consider a loan approval process.

A bank deploys three agents on the same file. One analyses financial history. One checks regulatory compliance. One evaluates risk.

Each is specialised. Each is correct within its domain.

But they do not always agree.

One signals approval.

Another flags regulatory concerns.

A third recommends waiting.

In a human team, this would trigger discussion, escalation, and eventual decision.

In a multi-agent system, the coordination must happen within the system itself.

A First Step: A2A

A standard known as A2A (Agent-to-Agent)—introduced by Google in April 2025 and now stewarded by the Linux Foundation—takes a first step in this direction.

Each agent publishes a capabilities statement:

• What it can do
• What it cannot
• How it prefers to receive tasks

Other agents can discover this, delegate accordingly, and negotiate handoffs.

It is not unlike a new team member introducing themselves—except that the introduction is machine-readable, and the negotiation happens in milliseconds.

Through A2A, each agent declares what it can do and how it prefers to work.

They pass tasks between each other. They escalate when needed. They signal disagreement explicitly.

The loan is neither approved nor denied multiple times.

It is routed.

Either to the next agent best suited to resolve the issue, or to a human when the question exceeds what any one system should decide.

It is coordination without a meeting.

What remains unresolved is the question of authority.

A protocol can indicate who spoke. It cannot determine who was right.

A protocol can indicate who spoke. It cannot determine who was right.

In human teams, that question is resolved through hierarchy, trust built over time, or the intervention of someone with enough standing to break the deadlock.

In multi-agent systems, that arbiter must still be defined.

In most organisations today, it remains a human—at least for decisions that matter.

The more interesting design question is not how to remove that human from the loop, but how to make their involvement precise.

To call them in at the moment where judgment is required—and not have them constantly manage what the system could have resolved on its own.

What We Choose to Standardise

The trajectory from the previous article becomes clearer here.

If behaviour becomes the basis of selection, and assessment becomes the basis of trust, then standardisation becomes the basis of scale.

What we standardise shapes what we can compose.

What we assess shapes what we allow.

And what we allow—quietly, one agent at a time—becomes the intelligence the organisation is willing to scale.

The firms that will navigate this transition well are not necessarily those with the most agents or the fastest deployments.

What we standardise shapes what we can compose.

They are those that have thought clearly about what they want to become—and have built the governance to hold that shape as the workforce around them changes.

That question does not remain theoretical for long.

It shows up in economics—cost structures, margins, speed, and risk.

In the final article of this series, we turn to those consequences.

Dig Deeper

Previously in The Future of Work and Agentic AI series

Part 3: When Intelligence Is Taken for Granted, Behaviour Is What Matters

Earlier parts:

Part 1: The Knowledge Divide: How AI Will Reorganise Work, Margins and Power

Part 2: When Work Runs on Two Minds

Next in the series

The final part examines the economic consequences of agentic organisations: margins, coordination costs, organisational design, and the changing structure of work itself.