The same model, run through two harnesses, costs 32x more in one.

The shift: from Pick the best model to Own the metered control plane

The coding-agent market stopped competing on model quality and started competing on whose tenant owns the meter, the audit log, and the kill switch.

From April to June 2026 the frontier models converged, parallel agents became substrate, and the open agent-skills standard made rules, skills, and MCPs portable across 30-plus tools, so the leaderboard score stopped picking a winner. What did not commoditize is the layer above the model: the control plane that meters spend, enforces policy at runtime, and can turn an agent off. GitHub Copilot flipping from flat seats to token-metered AI Credits on June 1, and Claude automation moving to full API rates on June 15, are the same event seen twice. The harness became a metered utility with shell access to a laptop full of credentials, and the renewal turned into a model-spend-plus-governance contract priced by finance, not a developer-tools line.

Early April artifacts

• Pick-a-favorite The harness market exited its pick-a-favorite phase. The harness market had a very loud week. Here's the part that changes your Q2 math.
• Model = risk Risk registers named the model and skipped the harness. The Harness Tax: A Boardroom Memo on Why the Wrapper Decides the Security Score

Mid June artifacts

• June 1 meter Copilot left flat seats for token-metered AI Credits. The coding-agent adoption number your board loves stops being free on June 1
• June 15 meter Claude automation moved to full API rates on a separate credit. Your AI coding agent's automation just got its own meter. What breaks on June 16?

Data signals

• +162% (2.6x) - Agentic-coding adoption more than doubled in two months while per-engineer cost rose on the same curve. Adoption ran from 32 to 84 percent in two months, so the June 1 meter turned that breadth straight into a variable bill the board now has to ration.
• 26 points - The same model, wrapped in two harnesses, swings 26 points on functional correctness. Wrapper-induced variance now exceeds typical model-to-model variance, so a risk register that names only the model has stopped describing the system it governs.
• 7 of 10 - Seven of the ten security posts trace to a named failure, breach, or unverified-defense class. No post argues a harness is securely solved, so treat the harness as an open attack surface and verify controls rather than buy more tooling.

Unresolved tensions

• Faster authoring versus scarce trust: At high AI adoption, task throughput per developer rose 33.7 percent as agents made authoring nearly free (Faros AI). vs The same high-adoption teams spent 91 percent longer in review as code volume climbed (Faros AI).. Read together, the two Faros AI numbers are one system, not a conflict: agents raise authoring throughput and shift the load to review, so the gain is absorbed by a longer verification tail. The contradiction is only apparent if a board reads the throughput number alone. Evidence: harness-cost-got-honest-productivity-number-didnt, harness-ai-code-review-tools-bottleneck.
• Cheaper tokens versus flat productivity: Vendor price cuts and metered billing make the cost side of ROI fall and look like a productivity win. vs Only 31 percent of enterprise AI spend ties to a business outcome, so a falling denominator on a flat numerator improves the ratio with nothing behind it.. A board can hear a productivity gain that is only a vendor price move, which is why the durable metric had to become verified merged output over cost rather than raw throughput. Evidence: harness-roi-number-cheaper-not-faster, harness-cost-got-honest-productivity-number-didnt.
• Portable primitives versus control-plane lock-in: Rules, skills, MCPs, and subagents are now open and portable across more than 30 tools. vs The orchestration glue, the meter, and the governance plane are vendor-specific and where the lock-in now lives.. Renewal math that treats the harness as one switchable line gets both the cost of staying and the cost of leaving wrong, because the standards moved down a tier while the lock-in moved up one. Evidence: harness-portability-audit-before-june-renewal, harness-one-or-portfolio-decision.
• Bottoms-up love versus admin-tenant control: Engineers picked the harness they preferred when admin-tenant features did not yet exist. vs Microsoft revoked its own engineers' beloved Claude Code licenses in favor of Copilot CLI, and GitHub flipped the base model under every tenant with no per-developer opt-out.. The adoption motion that drove every early harness win is being overwritten by procurement and platform owners, so rollout playbooks built on developer enthusiasm are now mispriced. Evidence: harness-adoption-myth-microsoft-retired.

Sprint moves

• Name a verification owner with real review capacity: Reorgs are cutting the authoring side of engineering while the load has moved to verifying agent output, which now waits 4.6x longer for human review across 8.1 million pull requests. Background agents already ship PRs overnight, so the 8am triage queue is an unowned production station accumulating debt this sprint. This Monday: Open one org-chart box between senior IC and engineering manager, write a one-page charter naming context curation, verification design, and blast-radius ownership, and assign it by name in next Monday's staff meeting. When agents write the code, who owns verifying it?
• Set a per-engineer token ceiling before the meter bites: GitHub Copilot left flat seats for metered AI Credits on June 1 and Claude automation moved to full API rates on June 15, so every active engineer is now a variable cost. A single high-output engineer can burn $47,000 in tokens in a month, and Uber already capped its leaderboard near $36,000 per engineer per year. This Monday: Ship a per-engineer token-budget config with a soft alert at 70 percent and a hard cap, wired to SSO and the billing API, and put the September credit-cliff number in the next finance review. When coding-agent usage drops after the meter, read it as a cost signal, not an adoption failure
• Inventory and scope every harness OAuth and MCP grant: Coding agents run with shell access on laptops full of credentials, and this quarter the Vercel pivot through one over-permissioned OAuth grant and the GitHub internal-repo breach via a poisoned VS Code extension both hit that surface. Only 11 percent of production agents sit in the fortified category, with 83 percent of claimed defenses never independently verified. This Monday: Ship a CI gate that fails any merge granting a new OAuth scope or non-allowlisted MCP server, plus a written permission inventory signed by the CISO before the next sprint review. vercel-breach-coding-agents-oauth-door

What to watch

• Whether Microsoft swaps the default GitHub Copilot model to its own Project Polaris on the announced August schedule, ending the three-month fallback window. (Q3 2026): A hyperscaler replacing a frontier model under every tenant without per-developer opt-out confirms the model has become a swappable, vendor-controlled commodity inside the harness. Trigger: Project Polaris becomes the default model in GitHub Copilot Business and Enterprise tenants with the fallback expiring.
• The first published case of a metered coding-agent invoice forcing a hard per-engineer spending cap or a credit-cliff response after the June 1 and June 15 billing changes. (within 90 days): Uber already capped two tools near $36,000 per engineer per year and a single engineer ran a $47,000 month, so the meter converts wide adoption into a variable cost boards must ration. Trigger: A named company announces a per-engineer token ceiling or a September credit-cliff response to metered Anthropic or GitHub Copilot spend.
• Whether engineering orgs begin reporting a verified-defense share or a named verification-owner role on the board slide instead of an adoption percentage. (H2 2026): With only 11 percent of agents fortified, 83 percent of claimed defenses unverified, and AI code waiting 4.6x longer for review, the standing board number is shifting from breadth of access to verified value and verified control. Trigger: A named enterprise publishes a verification-rate or verified-control metric in place of an adoption-rate KPI on a board slide.