Every enterprise buying agentic AI today asks the same question first: which model. A better first question exists. Two organizations can run the identical frontier model and reach different ceilings on what their agents execute, because the graph sets the ceiling and the model runs inside it.
The question arrives on a schedule. Gartner expects 40% of enterprise applications to carry task-specific AI agents by the end of 2026, up from under 5% in 2025. Each of those integrations inherits a shape from whoever drew it.
The workflow that gets slower every time it grows
A mid-market insurer wires its claims-intake agent as a chain. An extraction agent reads the incoming documents and passes its output to a summarization agent, which passes its output to a drafting agent, which passes its output to a review agent. Each step waits for the one before it to finish. Each step receives the full accumulated document set, because nobody built a reason to trim it.
The chain works. It also gets slower every time the business adds a document type, a jurisdiction, or a policy rule, since every new requirement lengthens the same single line and every agent in that line then reads more tokens than the task in front of it requires.
The cost here is measurable. A Transformer's attention mechanism computes a relationship between every token and every other token in its context, so the compute cost of reading a sequence scales quadratically with its length: O(n²), in the shorthand engineers use for it. Feeding a flat, ever-growing document stack through one chain of agents slows the workflow combinatorially, on every hop.
Two shapes, two ceilings
Linear execution queues steps in a line: agent A finishes, then agent B starts, then agent C starts, even when B and C share nothing. It is the shape a person draws while describing a workflow out loud, one step at a time. It is also the shape that hits a hard ceiling the moment the business adds a step, because everything in it sits downstream of everything before it and nothing can run concurrently.
The preference for the line has been measured. In 1942, Abraham and Edith Luchins ran the water-jar experiments that named the Einstellung effect: participants trained on a long solution method kept applying it when a far shorter route was available, at a rate of 98%, and 97% of them failed a later problem the long method could not solve at all. An untrained control group did markedly better on both. Told plainly to look again, more than half of the trained group found the short solution at once. The constraint was attention, and attention is what a drawn map supplies.
Graph-structured execution replaces the line with an explicit map of which tasks are independent and which are sequential, then routes work accordingly. Two mechanisms carry the load:
Fan-out spawns independent tasks concurrently across isolated agent sandboxes. Where the first task's output changes nothing about what the second task must do, both start at once. A claims intake workflow that fans out document classification, policy lookup, and prior-claim retrieval finishes all three in the time the slowest of them takes, and the other two cost nothing on the clock.
Tiering routes each node in the graph to the model tier its task requires. Administrative and extraction work — reading a form, classifying a document type, pulling a field — runs on small, efficient models built for exactly that job. The graph reserves its frontier, heavy-reasoning model for the judgment nodes at the end: the step where the organization needs synthesis, risk assessment, and language a human will read and rely on. An enterprise that routes every node through its most expensive model pays frontier-model compute to have a small model's job done slowly.
Both mechanisms stay out of reach for a workflow built as a line. A chain holds no independent tasks to fan out, since everything in it was forced into sequence, and it carries no tiering logic, since every link was built to hand off to the next link's full capability.
What the graph decides
Manta Graph™ names the shape before any controller executes against it: the structural, relational, and contextual map of what a given workflow requires. Which tasks are independent. Which are sequential. Which data each node has cause to see. Which sources feed which decision. That map turns "add an agent" into an architectural decision with a stated place in the graph and a stated reason for every field it touches.
This is the operational fact underneath the phrase agent topology: the graph determines what can execute, in what order, and under what constraint. An organization that skips drawing that graph has still made a decision about its execution topology. It has taken the straight line by default, with the ceiling that comes attached.
Across every governed workflow we have mapped, the pattern holds: the ceiling enterprises hit first belongs to the architecture, and it lifts when somebody draws the graph.
The graph decides what can execute. What decides which of those executions may proceed without a human standing over them is a separate question with a separate answer, and it is the subject of Part 2.
GlobalizeWe builds governed execution infrastructure for enterprise agentic work — Manta Graph™ to map the estate, GW Slate™ to control the run, Receipt Rail™ to prove it. Human-at-the-control, cryptographically attributable, built onto the systems an enterprise already runs.
Continue to Part 2: Agent Topology: Capability Buys Speed, Architecture Buys Authority