Models Swap in 48 Hours. The Harness Does Not

Model switching cost collapsed to days this week. Grok 4.6 launched on August 12, 2026 and was live inside Cursor the same day; within two days, GitHub announced it rolling out across GitHub Copilot’s eight named surfaces. This post makes an argument from that speed: the durable switching cost in a coding-agent stack no longer sits in the model — it sits in the harness around it. That is analysis, not a settled fact, and it is treated that way throughout.

The stakes are practical, not academic. Teams are signing contracts for coding agents right now, and most procurement conversations still center on which model a vendor offers. If frontier models reach rival harnesses within days of launch, exclusive model access is a wasting asset — and the questions worth negotiating shift to the layers that do not move: configuration, permissions, session history, integrations, and, as of this week, repository hosting.

What follows is the dated evidence chain, the argument built on it, the strongest counter-evidence we could find — stated plainly, not buried — and a procurement-ready map of which layers of an agent stack are genuinely portable at the time of writing and which are not.

No single vendor announcement contains this post’s argument. It is assembled from six dated events spanning August 6 through August 17, 2026, checked against five vendor primaries we fetched directly — with the August 6 spec publication reaching us through GitHub’s own August 12 changelog, which recounts it, rather than a primary we fetched ourselves. Two of them — Cursor’s Origin code-hosting beta and the Agent Plugins 1.0 GA — get full standalone treatments; this post uses them as links in a chain.

None of these vendors published this timeline. Each announced its own event in isolation; the cross-dated chain above required pulling four different company blogs and dating each entry against a vendor’s own published record, with the August 6 entry dated from a later recount rather than its own page. That assembly — and everything argued from it — is where this post goes beyond the record, and we flag it every time it does.

Start with the fastest link in the chain. xAI’s launch post for Grok 4.6, published Wednesday, August 12, states it directly: “Grok 4.6 is available today in Cursor and Grok Build.” Cursor’s own blog post the same day corroborates it independently. Day zero availability in a major third-party harness, confirmed by both sides.

By Friday, August 14, GitHub’s changelog announced the model “now rolling out in GitHub Copilot” — across Visual Studio Code, Visual Studio, Copilot CLI, the Copilot cloud agent, the Copilot app, JetBrains, Xcode, and Eclipse, on Pro, Pro+, Max, Business, and Enterprise plans. Two caveats from the same primary keep the claim honest: the rollout is explicitly gradual, so August 14 is the announcement date rather than a guarantee every seat had access that instant; and for Business and Enterprise plans, the Grok 4.6 policy is off by default until an administrator enables it. xAI also ran a first-week 2x included-usage promotion — a vendor-stated promo scoped to Cursor and Grok Build only, not Copilot.

The interesting comparison is between the two integrations, not within either one. Cursor’s day-zero availability has a plausible business explanation we examine in section 04. GitHub and Microsoft have no comparable corporate relationship to xAI, which arguably makes the two-day Copilot turnaround the more architecturally telling data point: a frontier model landing in a competitor-adjacent harness, across eight surfaces, within two days of existing publicly. Whatever else is true, “which vendor has which model” is now a question whose answer changes in days.

Here is the thesis, stated as what it is — our synthesis of the record above, not a fact any vendor has published. If a frontier model reaches rival harnesses within two days of launch, then exclusive model access cannot be a durable competitive advantage for any harness vendor, and it cannot be a durable switching cost for any buyer. What remains durable is the harness itself: the accumulated state that does not travel when you change tools.

Look at what that state actually is in a working coding-agent setup. Project configuration and instruction files tuned over months. Permission rules and allowlists that encode what your team trusts an agent to do. Hooks, custom skills, and plugin configuration. Session history and the context an agent builds up about your codebase. Integrations with your issue tracker, CI, and review flow. And as of this week, with Cursor’s Origin beta, potentially the repository itself — though Cursor’s own release copy keeps GitHub “the source of truth” for synced repos and explicitly defers agent-native features to a later release.

The vendors’ own behavior this week is consistent with that reading. OpenAI’s August 11 release did not ship a model — it shipped an importer, reading “instructions, settings, skills, plugins, projects, and recent work” from Claude Code, Claude Cowork, and Cursor into its own desktop app, with a matching /import command in the Codex CLI. You build an on-ramp for harness state when you believe harness state is what keeps users where they are. Cursor moving into code hosting the same week admits a parallel reading: the harness vendor closest to the repository is reaching for the stickiest layer available. We argued the complementary case — that the model layer deserves deliberate second-sourcing — in our vendor-resilience playbook; this week’s record suggests the model layer is becoming cheap to re-source almost by default, while the single-sourcing risk migrates up the stack.

Projecting forward: if this holds, harness vendors will compete less on model exclusives and more on accumulated state — richer project memory, deeper repo integration, agent-native features that live where your code lives. That is not inherently sinister; it is where the genuine product value is. But it changes what a buyer should scrutinize, because state that makes a tool valuable is the same state that makes it hard to leave. The rest of this post tests that argument and then turns it into questions you can ask before signing.

An argument you cannot disagree with is a take. Three pieces of counter-evidence deserve the floor — each grounded in the same week’s record, and each strong enough that a reasonable reader could finish this post unconvinced.

Counter one: the week’s biggest interoperability move cuts against the thesis. Agent Plugins 1.0 — GA on August 12 — exists specifically to make harness-level assets portable. GitHub’s own copy: “You can now build a plugin once and use it across all compatible agent clients.” If the harness were becoming the moat, the last thing you would expect is the harness vendors themselves co-publishing a spec that dismantles part of it. The honest scoping: thepackaging layer — skills and MCP configuration — is heading toward portable, and the thesis survives only where it is scoped to the state layer: context, history, permissions, integrations, repo hosting. Both readings come from the same primary source, and both are defensible.

Counter two: a two-day integration may say more about business development than architecture. Cursor’s parent, Anysphere, had been under a signed acquisition agreement with SpaceX since June 16, and the deal closed on August 14 — three days before Origin’s beta opened, and the same corporate family as xAI since their February merger. Grok 4.6 appearing in Cursor on day zero is unsurprising with a pending-merger relationship in view; it is weak evidence that any model can reach any harness that fast as a general architectural property. The admin-gated, explicitly gradual Copilot rollout — from a company with no such relationship to xAI — is the better evidence, and it took two days and still requires an enterprise admin to flip a policy that defaults to off.

Counter three: model quality differences are real, and a harness cannot paper over them. xAI’s own launch-post comparison table — vendor-compiled, with Grok’s row self-reported and competitor rows drawn from those vendors’ published system cards per xAI’s own caption — shows real spread on identical benchmarks, whichever harness any of these models runs in.

That is a 7.1-point spread between the top and bottom rows on a coding-specific benchmark — and on CursorBench v3.2, the same table has the ordering shuffled: Fable 5 Max at 70.5%, Grok 4.6 High at 69.9%, GPT-5.6 Sol Max at 67.2%, a 3.3-point spread with a different leader. Treat none of this as a settled ranking — it is one vendor’s compilation — but it is a directional signal that models are not interchangeable in outcome even when they become interchangeable in access. If model quality decides your workload, two-day availability does not make the choice for you.

Because the spec is both evidence for and against this post’s argument, it is worth being precise about what it covers. A spec-compliant plugin is a folder: a manifest, an optional skills directory, and an MCP server configuration file. GitHub’s GA changelog describes migration for existing plugins as mostly manifest work, and states that pre-1.0 Copilot plugins “remain supported, with no migration required.” We covered the spec’s format in depth when it was published in early August — at which point our post described it as still marked Working Draft. Four days after that coverage, GitHub’s dated August 12 changelog announced general availability across VS Code, Copilot CLI, the Copilot SDK, and the Copilot app, on all Copilot plans. That status change is real and citable; the precise date the spec’s own status field flipped is not something we could locate, so we date the GA to the changelog and nothing else.

No vendor publishes a “here is what will not come with you” table for its own product. Assembling one from this week’s record is the most useful thing this post can do. Every evidence cell below points at a dated vendor primary discussed above; the recommendation column is our analysis.

One more layer travels with the harness that this table cannot capture in a row: the data terms. Retention, training rights, and enterprise carve-outs attach to the harness vendor you sign with, not to the model you happen to route through it — and they differ more than most buyers expect. Our census of coding-agent data terms, published alongside this post, walks that layer vendor by vendor.

If the argument in section 03 is even half right, the negotiating leverage in an agent-stack purchase sits in different places than most 2025-era contracts assumed. Here is how we would walk into that conversation, layer by layer.

These questions are answerable in a procurement cycle, and the asymmetry favors whoever asks early: every layer is easiest to negotiate before state accumulates in it. This is the evaluation work our AI transformation engagements run for teams adopting agent stacks — and for teams building the routing layer underneath, our gateway architecture reference covers the infrastructure one level down, where multi-model portability is implemented rather than argued about.