Early access

An ontology describes the rules. A harness makes them stick.

OntologicalHarness is the governed scaffold an agent runs inside. It gives the agent meaning, constraints, authority and a path to act, so agents and humans work under one framework with full auditability, learning and insight.

Same rules for agents and humans
Same approvals
Same audit trail

Request early access See the platform

The problem and what it solves

Enable your AI agents the way you enable your people

You onboard a person with a role, clear rules, approvals and a manager who can see what they did. Most agents get a prompt and a credential. That leaves two problems.

The problem

Agents and people run under different rules

Agents act with broad access and no shared definition of the business. People work under policy and approvals. Two rulebooks means gaps nobody can audit, and nothing either one does teaches the other.

What it solves

One framework for both

The harness gives agents and humans the same meaning, rules, roles and approvals. Every action is checked and logged the same way, so you get full auditability, a record to learn from, and insight into what is working.

Auditability

Every action on the record

Who or what acted, which rule applied, who approved it.

Learning

Overrides become better rules

Near misses and overrides turn into proposed changes a steward reviews.

Insight

See how work really runs

Compare agent and human outcomes under the same rules.

One rulebook

Agents and humans follow the same rules

An agent does not get a side door. It holds a role, it hits the same checks, it needs the same approvals and it leaves the same record as a person doing the same job. The only difference between them is the role they hold.

Diagram: a human and an agent each hold a role and pass through one gate with the same rules, checks and approvals, producing the same outcome and an audit record — One gate for every actor.

Same checks

One set of rules

A BLOCK rule stops a person and an agent alike. A WARN flags both. Nobody writes a second rulebook for software.

Same approvals

No self approval

If a spend over the limit needs a manager, it needs one whether the request came from a person or an agent.

Same record

One audit trail

Every request, check, approval and outcome is recorded the same way, so an auditor reads one story, not two.

The gap

Agents are moving faster than the controls around them

Most teams can build an agent. Far fewer can say what it is allowed to do, stop it when it should not, and show an auditor what happened.

21%

of 3,235 leaders surveyed by Deloitte (2026) reported a mature model for agent governance.

~80%

in the same research lacked clear boundaries, monitoring or audit trails for agent actions.

63%

could not enforce purpose limits on their agents, in a Kiteworks survey (n=225). It is a vendor survey, so read it as a signal, not a census.

What the harness gives an agent

Four things a model cannot supply for itself

Meaning

An ontology says what things are and how they relate, in terms the business uses.

Constraints

Rules live as data. BLOCK stops an action. WARN flags it. Each has an owner, a version and a reason.

Authority

Agents hold roles like people do. Revoke the role and the agent stops.

A path

Dry run, precondition checks, approval, execution, audit. One controlled route to change anything.

Agents and humans understand relationships through the ontology and ask in business terms through the semantic layer. To act, they request an action from the action engine, which checks rules as data, records to the audit trail and executes against governed data. The learning loop reads audit evidence and proposes rule changes. — Two paths. Understanding reads the ontology and semantic layer. Acting goes through the action engine, which checks the rules.

How it works

Observe, understand, decide, act, confirm, learn

The loop runs the same way whether a person or an agent is driving. Learning proposes rule changes and cites its evidence. A steward approves. The agent never approves its own change.