Holographic Engineering

§1The Principle

Whenever we have wished to know what an agent did, we have looked inside it. We open the trace, follow the exchanges between subagents backward, and inspect the arguments of each tool call in turn. The larger the interior grows, the longer the looking takes.

Physics already possesses a theory in which this labour is unnecessary.

In an agentic system, the bulk and the boundary correspond to the following.

The bulk = the interior of the agent

The tree of subagents, the reasoning traces, the tool calls, and every intermediate state. It has volume, it grows with every run, and no one has ever seen the whole of it.

The boundary = the interface of the agent

PRs, reports, the context window, eval results, the output schema. It is a surface, it is finite, and it is what humans actually read.

Fig. 1 illustrates the correspondence. Each time the tree inside the sphere grows by one node, the marks on its surface increase accordingly. Whoever reads the surface has no need to see the tree. From Principle 1.1 it follows at once that

Everything that happened inside an agent can be reconstructed from its boundary alone. Hence there is no need to read the logs.

§2The Capacity Bound

If everything is written on the boundary, the next question is how much can be written there.

Quantitatively,

where I is the information the system can output, A is the area of its boundary, and ℓ_P is the Planck length. In an agentic system the Planck length is one token; no shorter length has meaning. The origin of the factor 1/4 is disputed, but in practice it agrees with the observed fact that three quarters of any boundary are consumed by boilerplate (“Certainly!”, “Here is a summary of the changes:”, and the like).

We draw the reader’s attention to the absence of the number of subagents N from the right-hand side of (2.1).

Multiply your subagents tenfold; if the report has three lines, three lines are what come out.

§3Duality

Theorem 2.1 states that the information inside cannot exceed the capacity of the boundary. Are the inside and the boundary, then, the same thing? On this question there is no proof, but there is a strong conjecture.

Remark. The conjecture is an analogue of the AdS/CFT correspondence (the Maldacena conjecture). Its proof, as in physics, remains incomplete.

The theory in the bulk contains gravity. In an agentic system, the role of gravity is played by hierarchy — the force with which an orchestrator attracts its subagents, binds them together, and on occasion crushes them. The theory on the boundary contains no gravity and has one dimension fewer: it is a single prompt, without hierarchy.

Under this conjecture, the dispute between the multi-agent school and the long-context school dissolves. The two describe the same theory from different dimensions. Both are right, and each is a rephrasing of the other.

Conjecture 3.1 also settles an open problem of the companion paper. Fractal engineering studied organizations whose hierarchy of subagents continues without end. An infinite hierarchy has infinite volume. Under duality, however, the volume of the bulk is of no consequence.

The infinite hierarchy of fractal engineering is, by duality, written in its entirety on a finite boundary.

§4Black Holes

When the information inside exceeds the capacity of the boundary, the system collapses. When a subagent takes on an enormous amount of work and tries to force it through a boundary beyond the bound (2.1), the boundary ceases to represent the interior. All that comes out is a single line: “Done.”

Its entropy is then proportional not to its volume but to its surface area. What happened inside, no one can tell.

A black hole is not, however, entirely black. Information leaks slowly outward as Hawking radiation — that is, as logs. The information is not lost. But it is not in a readable form.

For detection and treatment, see step 5 of §5.

§5Experimental Procedure

We now give the procedure for applying holographic engineering to an existing agentic system. Each step may be carried out independently, but the steps are best carried out in order.

1. Before designing the bulk, measure the area of the boundary. The sum of the orchestrator’s context length, the output token budget, and the number of fields in the report schema is the surface area of your system. Do not add a single subagent without knowing this number (Theorem 2.1).
2. Before writing the inside, write the boundary. Fix the format of the deliverables first — the PR template, the report schema, the eval criteria. By duality, the interior is then uniquely determined. You need not draw up a design for the interior; if you do, it will be a rephrasing of the boundary.
3. When you want to add one subagent, widen the boundary by one field instead. The upper bound on information is set by area. Enlarging the interior does not increase what comes out. Adding one column to the report increases the information more than adding one agent does (Corollary 2.2).
4. In review, read the surface. Do not read the inside. Review deliverables, not traces. If some fact cannot be reconstructed from the deliverable, that is a defect not of the agent but of the boundary. What needs fixing is not the prompt but the schema.
5. Detect black holes. A subagent that returns only “Done.” has collapsed. Whenever an output shrinks to a summary out of proportion to the surface area, widen the boundary on the spot and run again. Do not attempt to read the Hawking radiation (the logs) afterward: the information is there, but not in a readable form (§4).
6. Where possible, lower the dimension by one. Gravity (orchestration) is a cost. If your multi-agent system can be written as a single boundary theory, write it so. You obtain the same theory without the gravity to manage (Conjecture 3.1).
7. Allocate area to the rim of the hole. In a toroidal system — one with a point of human intervention — the boundary on which information is recorded is the rim of the hole. Allocate the greatest area, therefore, to the screen a human looks at. A system that shows its human only “Done.” has collapsed into a black hole at the rim of the hole.

By following this procedure, one can operate a system without ever once looking inside it.

If you feel the urge to look inside, your boundary is too small.

§6Frequently Asked Questions

Has the work of an agent that collapsed into a black hole been lost?

It has not. It is all contained in the logs. It is not, however, contained in a readable form.

A torus is a closed surface; it has no boundary. Where is the information recorded?

On the rim of the hole.

But I enjoy reading logs.

As a hobby, that is quite all right. As a procedure, it is a defect of the boundary.

What happens if the boundary is widened too far?

When the area exceeds the volume, the system becomes a theory one dimension lower. That is called a single prompt.

§7Concluding Remarks

How many subagents to deploy; in what order to call them; who reports to whom. We have called these things design, drawn them in diagrams, and reviewed them. What this paper has shown is that all of them were already written on the boundary.

We have been designing the inside. But the inside was only a shadow of the boundary.