The control plane is fractal

A common prediction says AI agents will replace apps. Rather than open several applications and move data between them, users will delegate an outcome.

What it gets wrong is the implied architecture. There will not be one agent above every application, comprehending every domain. Real systems do not flatten as they mature. They divide responsibility and delegate.

The agent economy will have many control planes, nested inside one another. Each layer orchestrates the systems beneath it and is a tool to the system above. From one direction it is the control plane. From another it is the destination. That is the architecture we built Pivot around.

Control depends on where you are standing

A control plane decides how work is decomposed, where state is held and when a human intervenes.

Legora describes its agent as the execution engine of an agentic operating system. Harvey is pursuing a similar position through agents, Vaults and shared matter Spaces, and inside Microsoft 365 Copilot, where Microsoft owns the surface. Harvey is therefore both a control plane and a destination. Control is relative.

General work agent (Copilot, OpenAI)

↓

Legal control plane (Harvey, Legora)

↓

Domain control plane (Pivot)

↓

Documents, evidence, valuations

A legal matter is not a prompt context

The meaning of a financial remedy matter does not reside in its documents. It resides in the relationships between parties, assets, liabilities, valuations, assertions and evidence. One party may describe an account one way while a statement identifies it differently. An asset may be alleged but not evidenced. A liability may be disclosed, disputed or duplicated. None of it survives collapse into retrieved passages of text.

Pivot holds the matter as structured state. Parties, assertions, evidence, conflicts, provenance and disclosure gaps are distinct objects with explicit relationships. Models can classify documents and extract candidates, but they do not get to decide silently that two records describe the same asset, or that an allegation has become fact. That is the work of the domain control plane.

What happens when another agent calls Pivot

A lawyer asks an upstream agent to prepare the asset schedule, identify unresolved discrepancies and show the evidence behind each figure.

The upstream agent should not improvise that from a large context window. It should delegate. Pivot resolves the request against the structured matter: which accounts refer to the same asset, which values are current and which superseded, whether a figure is asserted, evidenced or disputed, which source supports each entry, where disclosure is missing, which conflicts need judgement. It returns a structured result with provenance and review status.

An agent destination is more than an API

Making a product callable does not make it agent native. A dependable destination needs typed operations, machine readable schemas, scoped permissions and explicit approval boundaries. Above all, the semantic contract must survive the journey between systems. A disclosure gap reported by Pivot must not become an allegation in the next model. Interoperability without semantics only moves ambiguity faster.

The interface becomes a cockpit

An agent native interface exists for inspection, correction and approval: how the matter is represented, where records were joined, where the system refused to resolve an ambiguity, and how to correct it without losing history. Some users enter through Pivot, others through a platform above. Both reach the same state.

The specialist moat is the delegation boundary

If a product is a prompt, a generic workflow and a chat interface, an upstream plane reproduces most of its value. That changes when the specialist system owns persistent domain state, resolution rules, provenance, validation and the line between automation and judgement. An upstream agent may copy an interface. It cannot reproduce conclusions without rebuilding the system that gives them meaning.

Ontologies can be copied, so the defence is continuous: better structure, better rules, better evidence on every output. It is also institutional, because the tools an agent may call are constrained by procurement, information security and matter level authority.

The objective is not to own every interaction, but the point beyond which a broader agent cannot proceed by improvisation. That point is the delegation boundary.

Pivot is both layers at once

To a general agent or legal platform, Pivot is the destination for work needing structured intelligence about a financial remedy matter. Within that matter, Pivot is the control plane, coordinating extraction, identity, evidence, conflicts, provenance and review.

The question is not whether a lawyer uses Pivot or a horizontal platform. It is what that platform must entrust to Pivot when the matter needs more than plausible text.