Most technology benchmarks have a half-life measured in quarters. The SWE time horizon — the standard metric for how long a task an AI coding agent can reliably complete without human intervention — apparently has a half-life measured in weeks.
On January 14, Ajeya Cotra (previously Open Philanthropy, now at METR, the organization that conducts autonomous replication and adaptation risk evaluations for frontier AI labs) published her annual prediction: AI coding agents would reach a ~24-hour SWE time horizon by end of 2026. That was a reasonable, defensible number that enterprise IT teams could actually build procurement strategy around.
On March 5 — 50 days later — she published an update: the forecast is now greater than 100 hours, and possibly unbounded. Her words: "For the first time, I don't see solid evidence against AI R&D automation this year."
Kevin Roose at the New York Times flagged it with four words: "I hope everyone is paying attention."
For ops leads and IT buyers at 20-to-50-person firms who spent Q4 2025 running rigorous AI coding tool evaluations, this is not a feel-good headline. It is a procurement problem.
A Forecast That Aged in Weeks, Not Months
The reason the Cotra update matters more than the average Twitter AI hype post is the source. METR doesn't make noise for visibility. They run evaluations that AI labs — including Anthropic, Google DeepMind, and OpenAI — submit to voluntarily before releasing frontier models. When their principal researcher updates a quantitative forecast by 4× in 50 days and flags the ceiling as potentially non-existent, that is not optimism. That is an evaluator losing grip on the upper bound.
The January prediction wasn't conservative. A 24-hour SWE time horizon meant an agent that could reliably take a specification, a codebase, and 24 hours of compute cycles and return a working feature or fix — without human correction mid-task. Enterprise vendors were starting to pitch this capability as 2026's differentiator. Several already are: Devin (Cognition), SWE-Agent (open source), GitHub Copilot Workspace, and Cursor's new background-task mode all reference some version of autonomous multi-hour coding as their roadmap.
If the real capability ceiling is now past 100 hours and climbing, the marketing timelines your vendor pitched you in November are wrong. Not because vendors lied — because the field moved under them too.
SWE Time Horizon, Defined
For anyone whose evaluation process didn't go this deep: SWE time horizon is an empirical measure of agent reliability over time. It is derived from SWE-bench (the standard coding agent evaluation suite from Princeton/University of Chicago) but extended to measure autonomous task duration rather than pass rates on discrete problems.
An agent with a 4-hour SWE time horizon can independently resolve GitHub issues in a typical production codebase 50%+ of the time within 4 hours, without requiring human intervention. A 24-hour agent means overnight-batch-style autonomous PRs become a reasonable expectation. A 100-hour agent starts looking like a junior contractor who takes a ticket on Monday and returns a working branch by Thursday — with no Slack messages in between.
The operational implication shifts at each threshold. At 4 hours, you're reviewing output. At 24 hours, you're writing better specs. At 100+ hours, you're restructuring what the ops lead and the dev team actually do.
The Vendor Contract Problem
The uncomfortable truth is that most SMB-scale AI tool contracts are written against a capability snapshot that's already stale. Annual SaaS agreements signed in October 2025 based on Cursor Pro ($40/month), GitHub Copilot Business ($19/user/month), or an early Devin seat ($500/month) were scoped to a world where multi-hour autonomous coding was a roadmap item, not a shipping feature.
The tools will keep improving. That's the vendor argument. But "keeps improving" is not a performance SLA. And when the improvement curve moves faster than your renewal cycle, the evaluation methodology that justified the budget in Q4 doesn't apply to the capability you're actually running in Q2.
Three specific failure modes show up at firms in the 20-50 employee range:
Mismatched supervision workflows. A team that designed its review process for a 4-hour agent — human checks every PR, short feedback loops — is over-supervising a 100-hour agent. The overhead cost stays constant; the agent's autonomous output multiplies. That's a process debt accumulating weekly.
Underpriced tool seats. If a $40/month Cursor seat is now capable of replacing 8 hours of contractor time per day instead of 2, and your pricing was set at the lower capability level, you're paying commodity prices for what may now justify a different cost model entirely. Vendors will notice this before IT buyers do.
Benchmark drift in hiring. If your team evaluated candidates against "works with AI coding tools" as a skill in November, and the SWE time horizon has since tripled, the collaboration pattern you hired for may already be the wrong one.
Running Your Own Evals Before the Next Revision
The standard advice — "run your own internal benchmark before making tool decisions" — has always been right. The Cotra update adds urgency to the cadence. A benchmark your team ran in September is not a valid data point in March.
Concretely, for an ops lead managing a mixed engineering team: pick one well-scoped backlog item per quarter and run it through your primary AI coding tool with minimal intervention. Track wall-clock time to merge-ready, number of human interventions required, and whether the output required significant rework. Log it. Three data points will start to show the slope.
This is not glamorous. But it is the only way to know whether the tool you're paying for today reflects the capability generation you evaluated, or whether the floor has moved again underneath you while the contract stayed flat.
Cotra's January prediction was already aggressive by 2025 standards. Her March revision suggests even that was anchored to a capability landscape that lasted less than two months. In a product category moving at that speed, the most expensive mistake isn't picking the wrong vendor. It's trusting an evaluation you ran six months ago.