The Harness Is the Product, Not the Agent
Frontier models are converging and their prices are collapsing — the model is the commodity. The durable advantage is the harness around it: the context engine, feedback loops, and guardrails that make any model ship mergeable work.
Bottom line: model quality is getting cheaper and more uniform every quarter, so it cannot be a durable advantage. Whatever separates the teams that ship from the teams that thrash lives outside the model, in the system wrapped around it.
You are not buying a better model. You are renting one, and so is everyone you compete with.
That reframes the whole race. If the model is a rental, it cannot be your moat, because your competitor signs the same lease next quarter at a lower price. The model is the commodity. The harness around it is the product: the context, tools, feedback loops, and guardrails that decide whether a rented model ships mergeable work or generates plausible noise. This post makes the case for treating harness engineering as the real investment, and it names three costs most teams are paying without measuring.
Why is the model becoming a commodity?#
The economics are moving in one direction, fast. According to Epoch AI (2025), the price of running a model at a fixed capability level has been falling roughly 40x per year for GPT-4-class performance. When quality gets that much cheaper this quickly, quality stops being scarce.
Capability is also clustering. Stanford HAI's AI Index 2026 reports that top labs now sit closely together on public leaderboards, while SWE-bench Verified scores climbed steeply over a single year. When the frontier bunches up and prices fall, competition shifts away from raw capability toward cost and reliability. Anyone can rent intelligence now; the scarce skill is reliably turning it into shipped software. That gap, not the model, is where the fight actually happens.
| Dimension | The model (rented) | The harness (owned) |
| Cost trend | Falling roughly 40x a year | Compounds as your context accumulates |
| Availability | Identical to every competitor | Built in-house, unique to you |
| Moat | None; re-leased cheaper each quarter | Durable; the edge you keep |
If models are converging, where does the advantage come from?#
The advantage comes from the harness, and the evidence is unusually concrete. Anthropic's engineering team (2025) found that refining tool descriptions alone, without touching the model, pushed performance to the top of SWE-bench Verified. Their guidance is blunt: design token-efficient tools so agents don't burn a finite context window on overhead.
The same model, given a better-engineered environment, jumps to the top of a hard benchmark. That is harness engineering: the deliberate design of the context, tools, feedback loops, and guardrails that surround a model and shape what it can actually accomplish. The model is rented, so its gains flow to everyone. The harness is built in-house, so its gains compound for you alone. Prompt tweaks are tactics; the harness is the system that makes those tactics repeatable across every agent, repo, and engineer.
What is the context tax?#
Most teams are paying a large, invisible bill to move context around. Consider the numbers from Anthropic's "Code execution with MCP" (2025): a workflow that costs about 150,000 tokens through direct tool calls drops to roughly 2,000 tokens when the agent uses code execution instead, a 98.7% reduction. Large intermediate results can add another 50,000 tokens each time they pass back through the model.
CONTEXT TAX: the recurring token-and-time cost of re-feeding context to agents across sessions and tools because the system doesn't retain what it already learned.
That 150,000-to-2,000 gap is the tax made visible. Every time an agent forgets a repo, re-reads a doc it saw yesterday, or shuttles a giant payload through the model to touch one field, you pay again. The tax scales with your team and your ambitions, and it never shows up as a line item. It hides inside slower runs, bigger bills, and agents that stall halfway through a task. Harness engineering exists largely to stop paying it twice.
Why does token yield fall as you add context?#
More context is not automatically more understanding, and past a point it hurts. Chroma's Context Rot research (2025) tested 18 frontier models and found that accuracy degrades as input length grows, even on simple retrieval tasks, and well before the models reach their advertised context-window limits. Stuffing the window isn't a strategy; it's a slow leak that spreads the model's attention thin.
TOKEN YIELD: the amount of useful output you get per token spent, which falls as context bloats and the model's attention spreads thin.
This is why "just give the agent everything" backfires. Dumping the entire repo, every ticket, and a month of Slack into the prompt lowers your token yield instead of raising it. The winning move is selection, not accumulation: give the model the few things that matter for this task, and leave the rest out. Deciding what to include and what to withhold is a core harness responsibility, and it is precisely the judgment a naive "bigger window" approach throws away.
Are you actually more productive, or just faster?#
Speed and productivity are not the same thing, and the data now separates them. DORA's 2025 report found that AI raises throughput but is negatively related to delivery stability, summarizing the effect as "AI amplifies what's already there." Meanwhile GitClear's 2025 research found copy-pasted lines rose from 8.3% to 12.3%, refactored lines fell from 24.1% to 9.5%, and code churn climbed from 3.1% to 5.7%.
CONTEXT-ADJUSTED PRODUCTIVITY: engineering velocity measured net of the context tax and the rework it creates, rather than raw output.
Raw velocity flatters you. It counts every generated line as progress and ignores the churn, the reverts, and the review backlog piling up behind them. Faster typing at the keyboard means nothing if the outer loop, review and integration, jams. The harness is what converts speed into shipped, mergeable work: it feeds review the right context, catches regressions early, and keeps stability from cratering as throughput climbs. For more on that bottleneck, see the outer loop.
Harness engineering: build the environment, not the wishlist#
Stop shopping for a smarter agent and start building the environment it runs in. The market signal is loud: Menlo Ventures (2025) reports enterprise generative-AI spend hit $37B in 2025, up 3.2x year over year, with coding the single largest application category at roughly $4.0B. That money is chasing shipped software, not demos.
The load-bearing component of any serious harness is a context engine.
CONTEXT ENGINE: the layer that synthesizes and reconciles context across code, PRs, discussions, tickets, and docs, then delivers the right slice to each agent on demand.
This is the context engine for engineering, and it is what pays down the context tax and lifts token yield at the same time. Andrei Antanovich, a software engineer at Waste Logics, describes the pattern in practice: "The first instruction in every agent project file is: before making any changes, gather context." That instruction is harness engineering in one line.
On the context-maturity ladder, harness engineering is L6, the rung where the harness becomes load-bearing rather than a hobby. Tools like Unblocked build the context engine so agents start every task with decision-grade context instead of guesswork.
Frequently asked questions#
Is harness engineering the same as prompt engineering?#
No, though people conflate them. Prompt engineering optimizes a single instruction to a model. Harness engineering designs the whole environment around the model: retrieval, tools, memory, feedback loops, and guardrails. Anthropic's finding that better tool descriptions (2025) alone topped SWE-bench Verified shows the harness, not the wording of one prompt, drives durable gains.
What's the difference between context tax and normal token cost?#
Normal token cost is the price of work you needed to do once. The context tax is the price of doing it repeatedly because nothing was retained. Anthropic's code execution (2025) example makes it vivid: the same workflow ran at about 150,000 tokens versus 2,000, a 98.7% difference. That 148,000-token gap is pure tax, paid for context the system kept re-fetching.
If models keep improving, why invest in the harness?#
Because improvement flows to everyone equally. Epoch AI shows capability getting roughly 40x cheaper per year, so any edge you rent, your competitor rents too. The harness is the part you own. It compounds privately, gets better as your context accumulates, and turns whatever model you happen to be renting this quarter into reliable output. The edge you rent resets every quarter; the system you build keeps compounding.
Where the durable advantage lives#
Here is the whole argument in two sentences. The model is rented and commoditizing; its price falls, its quality converges, and every edge you find in it is available to your competition next quarter. The harness is owned and compounding; it absorbs your context, pays down the context tax, protects token yield, and improves context-adjusted productivity in ways no one else can copy.
So build the thing that lasts. Treat harness engineering as a real discipline, and make its center of gravity a system that delivers decision-grade context to every agent, on every task, without being asked twice. Models will keep getting cheaper and more interchangeable. What separates a team that ships from one that thrashes is the harness it built around the rental. If you want a map of the rungs from ad hoc prompting to a full harness, the context-maturity framework is a useful place to start.
