The Harness Is the Product
As agents move from prototype to production, the engineering discipline around the model — state, durability, oversight — matters more than the model itself.
24 August 2026 — The gap between an agent demo and an agent that runs a business process is not intelligence. It is engineering, and this week the industry finally started naming it.
The discipline emerging is harness engineering: everything that surrounds the model and makes it dependable. Durable state persistence — so that if a workflow fails halfway, it resumes rather than restarting from nothing. Explicit state machines — so the agent's possible transitions are enumerated rather than emergent. Oversight gateways — so a human sees the decision before it becomes a commitment. None of it is glamorous. All of it is what separates a system you can leave running from a system you must watch.
The evidence that this matters commercially is in the labour market. A $350,000 salary band has opened up for forward-deployed engineers — people who can sit between a business process and a technical implementation and make the two agree. That premium exists because the work is genuinely hard and genuinely scarce, and because organisations have discovered that buying a capable model gets them roughly ten percent of the way.
Anthropic's Project Parka, also announced this week, automates meeting follow-through — capturing commitments and pushing them into the systems where they need to land. It is a good illustration of the principle: the value is not the transcription, it is the reliable path from a sentence spoken in a room to an action recorded in a system of record.
What this means if you run an SME. When you evaluate an agent, stop asking what model it uses. Ask instead what happens when it fails mid-task. Does it resume, or does it start over? Does it know what it already did? Can you see, after the fact, every step it took and why?
Those questions will predict your experience with the tool far better than any benchmark. A modest model in a well-built harness will outperform a brilliant model in a fragile one, every time.
Twelfth in a weekly series on the developments that matter to small and mid-sized organisations.