Insights

Prompt to graph engineering: How enterprises build reliable agents on frontier models

Dr. Adnan Masood, Chief AI Architect, UST

Anthropic reports that multi-agent systems consume roughly fifteen times the tokens of a chat interaction. Knowing whether your problem justifies that multiplier is now a core architecture decision, and it depends entirely on which of five layers your failures live at.

In the last couple of years, what AI engineers worked on when building with large language models changed multiple times. It started as a string of text. It became a budget for what the model can see, then an environment that lets the model act, then a rule for deciding when unattended work is finished, and finally a map of which specialized agent does what.

Every serious AI system now comprises five layers, whether designed explicitly or assembled implicitly. Rewording a prompt cannot give a model a fact it never learned. Better retrieval cannot run a test suite. This history matters to enterprise leaders because most AI programs are still funded, staffed, and measured as if the prompt were the whole product.

DIVIDER

The five layers, in plain terms

These concerns accumulate. A production agent holds all five at once.

Prompt

A prompt is the authored instruction that conditions a model for one request. An insurer classifying incoming claims into severity tiers with a system prompt and a handful of examples is operating at this layer. It is cheap, fast, and easy to debug. Its limit arrives the moment the model needs knowledge it lacks.

Context

Context engineering is the practice of deciding, on every call, which information the model sees: retrieved policy documents, prior conversation, customer records, tool results, and memory, all within a finite token budget. The term entered mainstream use in June 2025, and Anthropic formalized it that September. Bigger context windows did not solve the problem: Chroma tested 18 frontier models and found accuracy on some tasks falling more than 30 percent at tens of thousands of tokens, well below rated limits. A payments company grounding a service agent in current fee schedules and dispute rules is doing context engineering whether it uses the term or not.

Harness

A harness is everything around the model that lets it act: the tools it can call, the sandbox it runs in, the state it keeps between steps, the permissions that limit what it may touch, and the recovery logic that handles failure. Anthropic defines it as the system that enables a model to act as an agent, and notes that any agent evaluation measures the model and the harness together. A 2026 study raised a coding agent’s benchmark pass rate from 69.7 to 77.0 percent by changing only the tools and middleware around the model. A claims-adjustment agent that reads a policy, queries a core system, drafts a settlement, and routes anything above a threshold to a human is a harness problem before it is a model problem.

Loop

A loop is a harnessed run that repeats without a person in the seat until a verifier confirms the work is done. The verifier is the deliverable. Software teams have the easiest time here because compilers and test suites supply cheap ground truth, which is why overnight code modernization loops matured first. Without that kind of feedback, stop at the harness until a real verifier exists.

Graph

A graph is a topology of specialized agents with explicit routing between them. Anthropic’s production research system, a lead agent coordinating several subagents with isolated context, outperformed a single agent by 90.2 percent on its internal evaluation. The same report notes the system used roughly 15 times the tokens of a chat interaction. A compliance function that dispatches a regulatory change to parallel agents covering policy, controls, and customer communications, then synthesizes their output, is a graph. Most tightly coupled work, including most coding, is not.

DIVIDER

What the evidence says about cost

Every layer buys reliability with money, time, and complexity. Anthropic reports that ordinary agent interactions consume roughly four times the tokens of chat and multi-agent systems roughly fifteen times. Reliability compounds in the other direction: a 50-step process where each step succeeds 99 percent of the time completes successfully about 60 percent of the time. That arithmetic explains why mature agent programs spend disproportionately on verification, checkpoints, and recovery rather than on prompts.

DIVIDER

Why working close to the model builder matters

Much of the clearest guidance on this subject comes from Anthropic’s own engineering publications: the December 2024 taxonomy separating workflows from agents, the 2025 context engineering post, and the multi-agent research system write-up. Those documents describe how the people who train the model expect it to be operated. That expectation shows up in behavior, since models are increasingly post-trained against specific harness conventions.

UST is an Anthropic Global Premier Partner, which puts our engineering teams close to that guidance as it develops and gives our clients an implementation partner who has already worked through it on Claude in regulated production environments. Our practice has delivered agentic systems for healthcare payers, insurers, and financial services firms, where the harness and the verifier carry regulatory weight. Two accelerators come out of that work. UST-Eval gives clients an evaluation harness so that every layer above the prompt has a verifier before it goes live. ResponsibleRails supplies the governance controls, permissions, audit trails, and containment boundaries that enterprise risk teams require before an agent is allowed to act on real systems. A red-teaming toolkit tests agents before deployment, because an agent that can act has a blast radius, and you can measure it before production.

None of that replaces the judgment described above. It shortens the distance between deciding which layer a problem lives at and having a working, governed system at that layer.

DIVIDER

What to do next

Start by locating your current failures on the five layers. If the model says things badly, that is a prompt problem. If it does not know things, that is context. If it cannot act, or acts unsafely, that is the harness. If it cannot run unattended, that is the loop. If a single agent cannot cover the breadth, and only then, that is the graph.

Then build the evaluation set before anything else, because every layer above the first depends on it. Teams that do this well spend less, ship faster, and can show their boards exactly what reliability each dollar bought.

If you are planning an agentic program on Claude and want a partner who has already worked through these layers in regulated industries, our UST AI Practice would welcome the conversation.