AI agents hit a context wall

METR, the nonprofit that runs the standard benchmark for how much unsupervised work an AI agent can handle, published a new number in May 2026. The latest edition of that benchmark, called Time Horizon 1.1, found that frontier agents now succeed, half the time, on autonomous tasks spanning two to four hours. That figure has roughly doubled every seven months for six years running.

If you’ve been reading about growth loops, that number sounds like exactly what you were promised. The pitch going around growth and product circles is that loops, the compounding acquisition-to-retention cycles that outperform one-off marketing pushes, are shifting from something a team runs by hand to something an agent runs continuously. Ian Vanagas, who leads growth engineering at the product analytics company PostHog, has argued loops are the next frontier precisely because agents can now automate the iteration cycle itself (2026). Read the METR number next to that pitch and it looks like confirmation: agents can already handle multi-hour work, so hand them the loop and let it compound on its own.

That reading skips a detail that changes the whole plan.

The advantage doesn’t go to whoever automates their loop first. It goes to whoever maps precisely where the agent stops and a person has to take over, and keeps redrawing that line every time the model improves.

The detail is what METR’s two-to-four-hour ceiling actually measures. It’s built on “well-specified, low-context” work: tasks scoped the way you’d scope something for a new hire who’s never met your team. The moment a task needs cross-person interaction, institutional history, or a judgment call that resists a clean pass/fail check, success rates drop sharply. METR doesn’t treat that as a rough edge it expects to sand off soon. It treats it as structural: a durable gap between what agents can execute and the contextual judgment that real professional work runs on.

Now look at what a growth loop actually is. It’s cross-functional by definition, touching acquisition, activation, and retention at once. It’s multi-stakeholder: product, marketing, and engineering all have a hand on the wheel. And it depends on institutional knowledge, what your users actually want, how your business model makes money, where your competitors are exposed, that no one wrote into a spec. That’s not an edge case for the agent-autonomy ceiling. That’s the exact profile METR’s data says agents handle worst.

PostHog’s own production examples make the point better than any warning label could. When Vanagas’s team runs agents inside real loops, the agents aren’t running the loop cycle end to end. They’re doing bounded, mechanistic pieces inside it: executing one specific analysis, deploying one specific experiment, while a person still decides what the loop should test next and why. The successes that already exist are successes at narrow delegation, not at autonomy.

There’s a second constraint stacked on top of the first. A separate set of 2026 industry cost analyses puts project failure at 60% for AI initiatives that don’t yet have AI-ready data infrastructure. Even the narrow, well-scoped sub-tasks agents can technically handle still need an organization that can hand them clean, usable context, and most companies can’t do that yet. So the bottleneck isn’t only what agents can do. It’s whether your data pipes can feed them the parts they can do.

Put those together and the “infrastructure maturity wins” framing floating around growth circles turns out to be right, for a different reason than advertised. It isn’t that the company with the best infrastructure gets to full loop autonomy first. It’s that the company with the best infrastructure can see its own agent-human boundary most clearly, and keeps moving that boundary as the ceiling rises instead of guessing where it sits. Scope agent deployment tightly, one bounded slice of the loop at a time, and you get real gains. Hand the whole loop over because the trend line looks encouraging, and the multi-stakeholder step nobody scoped for is where it breaks.

None of this means the ceiling is fixed. METR’s own trend has doubled roughly every seven months for six years, and context tools like retrieval, shared memory, and the Model Context Protocol, the industry standard for connecting agents to tools and data, are actively trying to close the exact gap that causes the context-fragility problem. It’s possible that gap narrows faster than anyone planning around today’s loops expects.

That’s an open question, not a settled answer, and it’s worth asking honestly instead of assuming your way past it: as the ceiling keeps rising, does better context tooling actually dissolve the multi-stakeholder wall, or does context-fragility turn out to be a structural floor that keeps a person permanently in the loop on the parts of growth work that touch other humans? Nobody has the data to answer that yet. Until someone does, don’t plan around the trend line. Draw the map of your own loop today instead: which steps are bounded enough for an agent right now, and which ones still need a human who carries context the agent never had.