The Context Problem

Read alongside: The Oversight Illusion and What A Real Decision Record Looks Like. The full product architecture is written up in the AI Agent Accountability Framework case study.

Some AI failures do not look like failures until later.

The agent does not hallucinate. The logic is sound. The recommendation is technically valid.

And it is still wrong.

It is wrong because the agent does not know what the business knows: a supplier is under review, a customer relationship is sensitive, a cost freeze has just landed, a route is politically awkward, an authority threshold has changed.

The agent reasons correctly against the information it has.

The problem is the information it has is incomplete.

More Data Is Not The Same As More Context

The instinctive response is to connect the agent to more systems.

Give it the ERP. Give it the CRM. Give it contract data, communications, support logs, finance data and whatever else is lying around looking important.

This can help, but it does not solve the core problem.

Operational data tells the agent what exists. It does not reliably tell the agent what matters right now.

A supplier's delivery performance may look fine in the data. The business may still know that legal is reviewing that supplier and no new commitments should be made. A route may look efficient. The business may know a new logistics partner is being trialled and should be preferred. A procurement action may be within policy, but outside the current CFO's cost-freeze intent.

Facts are not enough.

Agents need structured business context.

What A Manifest Does

The governance manifest is my answer to this context gap.

It is a structured markdown document, version-controlled under git, that captures the context agents need to reason against: priorities, active risks, constraints, authority rules, escalation thresholds and operational sensitivities.

Not raw data.

Context.

The manifest should contain the kind of information a senior new hire would need before making decisions on behalf of the business.

What are we optimising for this quarter? Which risks are active? Which suppliers are qualified? Which decisions can be handled routinely, and which need director approval? What has changed this week that overrides the default policy?

That is the missing layer in many agentic systems.

Why Markdown Under Git

The manifest needs to be readable by humans and usable by agents.

That is why markdown matters. A procurement director, compliance lead or operations manager should be able to open the document and understand the rules shaping the agent's behaviour.

Git matters because context changes are governance events.

If a risk threshold changes, an approved supplier list is modified, or a new escalation rule is added, the organisation needs to know who changed it, when and why. Version control gives you history, diffs, review and rollback without inventing a fragile governance database ceremony.

The manifest becomes a living document with a defensible history.

The Feedback Loop

The manifest should improve as the human-AI system learns.

When a human overrides a proposal, that override is a signal. Sometimes the agent got the data wrong. Often, the agent missed context that should have been captured.

If the same type of override keeps appearing, the answer is not necessarily model retraining. It may be a manifest update.

The human updates the governance context. The agent reads better context next time. Future proposals become better calibrated.

That is a quieter kind of learning, but in enterprise settings it may be the more important one.

The Practical Move

More data may make an agent busier. Better context makes it safer and more useful.