Beyond the frontier model: the real economics of agentic workflows

For the past couple of years, enterprise AI strategy has defaulted to one move: connect to the biggest frontier model available and send it everything. For chat assistants and general coding help, that default is defensible.

Agentic workflows break the default, because they change the shape of the bill.

Where agent costs actually come from

A chatbot’s cost model is simple: one prompt in, one answer out. An agent’s isn’t. Give an agent a real task — optimize this supply-chain route, reconcile these accounts — and it plans an approach, calls a dozen tools, evaluates the responses, discovers a constraint it violated, and starts over. Then it does that again.

The expensive part of an agentic task is not the first answer. It’s everything after: the checking, the repairing, the re-verifying. Teams budgeting agent workloads on prompt-plus-output math consistently underestimate what they’ll spend, because the refinement loop — the very thing that makes agents useful — is the dominant cost, and it multiplies with task complexity.

Run that loop through a premium-priced frontier API and the economics collapse quickly. For a bounded, domain-specific task, paying frontier rates for every intermediate reasoning step is hiring a principal architect to format a spreadsheet: the work gets done, and the invoice makes no sense.

The open-weight shift

This is the economic logic pulling enterprises toward open-weight models — models whose weights you can run on your own infrastructure. The current generation is capable enough for well-scoped domain tasks, and running in-house changes the cost per iteration from a metered API bill to hardware you’ve already bought.

Cheap iteration is not a small win. It’s the whole game, because of how agents improve: an agent that can afford to explore a wider search space, test several candidate solutions, and verify its own work before answering will beat a smarter agent that can only afford one attempt. When iteration is nearly free, you stop rationing the agent’s thinking — and on bounded tasks, thorough usually beats brilliant.

Tune the harness, not the model

There’s a deeper point under the cost argument. The model sets the ceiling on output quality. The harness — the orchestration layer that controls prompts, tool definitions, retrieval, and working memory — determines how efficiently the agent reaches that ceiling.

Most teams over-invest in the ceiling and under-invest in the climb. A mid-sized open-weight model in a well-tuned harness, with sharp tool definitions, good retrieval, and rigorous evaluation loops, routinely outperforms a frontier model in a sloppy one — at a fraction of the run cost. The harness is engineering you own and improve. The model is a component you swap as the field moves.

Where frontier models still win

Honesty requires the other column. Frontier models remain the right call for open-ended reasoning, for tasks where the problem itself is poorly defined, for low-volume/high-stakes work where the marginal cost doesn’t matter, and as the benchmark you evaluate everything else against. The mistake isn’t using frontier models. It’s using them as the default for every step of every workflow, including the thousands of bounded refinement steps a cheaper engine handles just as well.

The practical rubric: bounded task, defined success criteria, high iteration volume — open-weight in your own runtime. Open-ended reasoning, undefined success, low volume — frontier. Most mature agent architectures use both, routed by task.

Where we stand

This is why we designed Cortex to be model-agnostic — in our architecture, the model is a component inside the boundary, not the boundary itself. The delivery guarantees come from the layer around it: plans, validation contracts, and evidence. When a better or cheaper engine appears, we swap the component and keep the guarantees. Our thesis on why that layer — not the model — is where the real bottleneck lives is in our Field Paper, Why the Bottleneck Moved From Writing Code to Delivering It.

Questions this paper answers

Why are agentic workflows more expensive than chatbots?

Because agents loop: they plan, call tools, evaluate results, and retry until constraints are met. Most token spend sits in that refinement cycle rather than the final answer, and it multiplies with task complexity — so prompt-plus-output budgeting consistently underestimates real cost.

When should an enterprise use an open-weight model instead of a frontier model?

For bounded, domain-specific, high-iteration tasks with defined success criteria — where cheap per-iteration cost lets the agent explore and verify thoroughly. Keep frontier models for open-ended reasoning, ill-defined problems, and low-volume, high-stakes work.

Does a smaller model mean worse agent results?

Not on bounded tasks. The model sets the quality ceiling, but the harness — prompts, tools, retrieval, memory, evals — determines how efficiently the agent reaches it. A well-tuned harness around a mid-sized model routinely beats a frontier model in a sloppy one.

What’s the hidden cost in AI agent deployments?

Two things: the refinement tokens nobody budgeted, and the human rework needed to make agent output production-ready. Both are harness and validation problems more than model problems — which is why the orchestration layer deserves the investment.