Back to Research
ResearchMarch 20268 min read

Agents Can Now Act. But They Still Cannot Be Trusted.

The agent stack is now technically complete. It is still missing the one layer that matters in regulated systems: decision governance.

The Stack Is Finally Here

Over the past twelve months, the building blocks of the agent economy have come together with surprising speed.

Intent layers such as MCP enable agents to discover capabilities and coordinate tasks. Payment protocols like x402 and MPP define how value moves between parties. The Open Wallet Standard introduces a unified vault for holding funds and signing transactions. Settlement layers, including blockchains and real-time payment rails, finalise economic activity.

For the first time, an autonomous agent can:

  • receive an instruction
  • determine a course of action
  • execute a transaction
  • complete settlement

without human intervention.

This is not incremental progress. It is a structural shift in how software interacts with the economy.

The Industry Is Solving the Wrong Problem

The dominant narrative is that the final missing piece was the wallet.

It was not.

Wallets solve custody. They do not solve control.

The real question is not:

Can an agent hold and move money?

It is:

Under what authority is the agent allowed to act, and how is that authority proven?

This distinction is subtle, but foundational.

Execution Is Not the Same as Permission

Today’s agent stack can execute actions with increasing sophistication. But execution alone is insufficient in any regulated or high-stakes context.

What is missing:

  • Authority context
    Who or what granted permission for this action?
  • Policy enforcement
    What constraints governed the decision at the time it was made?
  • Deterministic evidence
    Can the system reconstruct exactly what was known, decided, and executed?
  • Liability attribution
    When something goes wrong, who is responsible, and on what basis?

Without these, agents are operationally powerful but institutionally unusable.

The Failure Mode Is Already Clear

Most current implementations follow a familiar pattern:

The agent is given access to funds.
The agent is given a task.
The agent executes the task.

If the outcome is questioned, teams fall back on:

  • logs
  • prompts
  • partial system state
  • human interpretation

This is the same architectural flaw that has plagued financial systems for decades:

Governance is reconstructed after the fact, rather than enforced at the point of action.

As stakes increase, from small transactions to portfolio rebalancing, lending decisions, or treasury management, this model breaks.

The Missing Layer: Decision Governance

The agent stack is often described as four layers:

  • intent
  • payment
  • wallet
  • settlement

What is missing is a fifth, more fundamental layer:

Decision Governance

This layer sits above execution and answers three critical questions:

  1. Should this action be allowed?
  2. Under what constraints is it allowed?
  3. Can the decision be proven and replayed?

Without this layer, the rest of the stack cannot safely operate in financial systems, superannuation, or institutional environments.

From Software to Systems of Record

Traditional software systems optimise for:

  • functionality
  • performance
  • user experience

Agent systems must optimise for something else entirely:

Defensibility

In financial contexts, it is not enough for a system to be correct. It must be provably correct under scrutiny.

This requires:

  • append-only evidence models
  • explicit authority boundaries
  • policy-driven execution
  • deterministic replay of decisions

These are not features. They are architectural requirements.

Why This Matters Now

The convergence of agent frameworks and payment infrastructure means that agents will begin executing real economic activity at scale within the next 12 to 24 months.

The immediate second-order effects are predictable:

  • increased transaction velocity driven by automation
  • new forms of machine-to-machine commerce
  • rapid experimentation in consumer and enterprise workflows

But the harder problems will emerge just as quickly:

  • disputed transactions initiated by autonomous systems
  • regulatory challenges around machine-held balances
  • breakdown of accountability in multi-agent workflows

These are not edge cases. They are the natural consequences of removing humans from the execution loop without replacing the governance layer.

The Next Phase of the Market

As execution becomes commoditised, value will migrate.

Not to:

  • faster wallets
  • cheaper transactions
  • more capable agents

But to:

  • systems that can prove why actions occurred
  • systems that can enforce policy before execution
  • systems that can attribute responsibility with precision

In other words, to systems of control and trust.

The Turing Dynamics Thesis

Turing Dynamics is built around a different assumption:

Governance cannot be retrofitted onto agent systems. It must be embedded into their core.

Rather than treating compliance, audit, and control as external overlays, the system is designed such that:

  • every action passes through an explicit authority boundary
  • every decision is evaluated against executable policy
  • every outcome produces a verifiable evidence trail
  • every state can be deterministically reconstructed

This transforms the role of the system:

From:

a tool that executes instructions

To:

a system that governs execution itself

A New Primitive for the Agent Economy

If the Open Wallet Standard becomes the SMTP of agent payments, the next required primitive is not another protocol for moving money.

It is a protocol for verifying decisions.

A system where:

  • actions are inseparable from their authority context
  • execution is inseparable from policy validation
  • outcomes are inseparable from evidence

Only then can agents move from:

capable of acting

to:

trusted to act

Closing Thought

The industry has crossed an important threshold. Agents can now participate in the economy.

But participation is not the same as legitimacy.

Until systems can answer clearly, deterministically, and defensibly:

  • why an action occurred
  • under whose authority it occurred
  • within what constraints it was allowed

agents will remain powerful, but fundamentally untrusted.

The next phase of the agent economy will not be defined by what agents can do.

It will be defined by what they are allowed to do, and what can be proven about it.

Continue

Continue through the research and platform surfaces.

Read more research or move from the editorial layer into the platform thesis.