The scarce thing is no longer access to intelligence — it is the accumulated system around the model that makes intelligence yours. This layer is where that system lives, and where the action is.
The harness is everything wrapped around the model: memory, context, skills, workflows, evals, integrations, business rules. For an enterprise, that bundle is what Alex Karp calls its "alpha" — the proprietary context that makes a generic model perform like it works for you.
A harness that hands an enterprise its alpha creates very low, model-agnostic switching costs. The model underneath becomes a component; the harness holds the relationship, the context, and the compounding.
The boundary: this layer owns no weights. Its product is the membrane between users and models — routing, privacy, policy, context, tools, guardrails. Making the weights is the layer below; what end customers buy is the layer above; the membrane in between is what this page is about.
Hold the model fixed, change the scaffolding, and the benchmark score moves as much as a model-generation upgrade.
Hand-curated from the SWE-agent paper (Yang et al., NeurIPS
2024) and Anthropic's Claude 4 announcement; committed as
data/harness-uplift.json. Benchmark variants differ across
pairs — compare within a pair, not across pairs.
The sharpest recent datapoint: Intelligent Internet's Zenith — an adaptive, self-improving harness — took GPT-5.5 from fifth place under its native Codex harness to the top of the FrontierSWE leaderboard, ahead of frontier models running their own native harnesses. Same model, different harness. FrontierSWE scores by rank rather than percentage, so the chart above draws on the longer SWE-bench record, where same-model scaffolding deltas are published as percentages.
The harness captures roughly $0.10 today — the smallest slice on the strip. The bet is that this is the slice that compounds: switching costs deepen with every skill, eval, and workflow the harness accumulates.
Illustrative split — assumptions in Methodology.
Work inside a rented assistant and everything you create runs inside their product — their memory format, their pricing, their rules, their ceiling. You can export your content, but content is not the system: an export is not a runnable copy of your memory, tools, automations, and routing. When you leave, the system stays behind.
Own the layer around the model and the same decade compounds into a system you hold: memory as a database you can open, skills as files you can read and copy, automations you can inspect, profiles you can move. The model underneath is just a component — swapped at will for cost, reasoning, or policy, eventually fine-tuned on your own accumulated data.
"The scarce thing is no longer access to intelligence, it is the accumulated system around the model that makes intelligence yours."
The market just put a number on this. In July 2026, Stripe was reported to have entered exclusive talks to acquire OpenRouter at roughly $10 billion — about 7.7x the valuation set by its funding round ten weeks earlier (first reported by The Wall Street Journal and The Information on July 23–24; still unsigned as of early August). Gateways are structurally thin — a markup on someone else's compute that a competent team can clone in a quarter — so a price like that only makes sense as an option on the configuration layer. As "The Everything Router," a widely read analysis of the deal, put it: your model can be swapped in an afternoon, but your fifty OAuth grants, your permission scopes, your approval policies, and your audit trail cannot. That configuration adds a switching cost that models by themselves never had.
Agent-infrastructure investors make the same argument from the other end: value in the stack accrues to wherever the agent's decision and policy layer lives — spend limits, allowed services, per-transaction caps, governance controls. That layer is the harness. Guardrails are the product here; this is where trust gets manufactured, and where the credential sits when an agent goes to buy something on your behalf.
The labs are coming for this layer. OpenAI's Presence, launched in July 2026, sells routing, access control, pre-deployment simulation, and audit trails inside the model contract; Anthropic has made managed memory a platform default and moved tools server-side into its API. There is a future where the model's API eats the harness. Pushing the other way: every CTO is building a vendor-agnostic AI architecture, and enterprises resist lab lock-in — which is precisely the opening for neutral membranes, portable across models. The membrane claim survives that squeeze only with a qualifier attached: independent of any one model. A harness owned by a lab is distribution for that lab's weights; a neutral harness is the customer's own asset.
This site practices what the layer preaches: the ask-AI chat in the corner runs through a privacy-membrane route — Venice's end-to-end-encrypted serving tier.
Devin — the harness as a full software-engineering coworker.
Hermes Agent — a complete agent you run and own; MIT-licensed runtime, portable everything. Reportedly raising at a $1.5B valuation (July 2026).
Hands enterprises their data, workflows, evals, and business rules as an owned system.
Forward-deployed engineers embed a company's data, workflows, and business rules — its alpha — into operational AI systems, on whatever model sits underneath.
Context is the product — the editor as the accumulation point.
One API over hundreds of models, routed for cost and uptime. Stripe reportedly entered exclusive talks in July 2026 to buy it at ~$10B — ~7.7x its May round, unsigned as of this writing.
The privacy membrane between users and every model — the more models commoditize, the more the membrane is the product. Where OpenRouter routes for price and availability, Venice routes for privacy: E2EE, no-logging, token-aligned — verifiable non-custody and policy ownership across models it doesn't host. Cross-layer reach both directions: downward, its own open-weight E2EE GPU serving tier; upward, consumer products (private chat, Characters, Agentic Chat) on the Applications & Agents page — the membrane worn as a product.
Developer plumbing for teams building their own harness from parts.
Whoever owns the system around the model. Swapping a model takes an afternoon; your context, evals, permission scopes, and audit trail take years to accumulate, and they compound. Reliability itself gets manufactured in this layer — raw agents decay toward coin-flips over long tasks, and guardrails are what turn that into work someone will pay for. It's why I put Venice here: routing every model through a privacy membrane is a harness business, and the more models commoditize, the more the membrane is the product. The chat on this site runs on exactly that pattern, which felt like the honest way to make the argument.