For the past three years, the phrase “jagged frontier” has been one of the most useful mental models in AI strategy. The basic idea: AI is surprisingly good at some things and surprisingly poor at others, with no reliable pattern. A model might draft a strong legal brief but fumble a simple arithmetic check. It shapes how teams decide what to delegate and what to keep. It was in constant use.
The data, on reflection, was misread.
The way we were measuring was the problem
When researchers established the jagged frontier, they were mostly running single-shot, single-agent evaluations — ask the AI once, take the answer, score it. That setup removes every piece of organizational structure we normally apply to hard work: no ability to try something, recognize it failed, and try again; no parallel exploration of different approaches; no accumulated context across a multi-day effort; no one playing the role of critic.
Describing AI capabilities under those constraints is a bit like assessing a surgeon’s skill by asking them to operate in the dark with one hand behind their back. The result reflects the setup, not the intelligence.
The constraint was in how we asked the question. The jagged frontier was measuring “how good is AI at one-shot work with no scaffolding?” — and most real work isn’t one-shot, and most capable professionals don’t work alone without structure.
One result that reframes everything
Earlier this year, the CEO of Cursor announced something worth returning to. Cursor’s general-purpose coding harness — a system built to write software — ran for four days on a research-grade spectral graph theory problem (spectral graph theory is a branch of mathematics dealing with properties of graphs through the lens of linear algebra, well beyond software development). The problem came from unpublished work by academics at Stanford, MIT, and Berkeley.
No human guided it. No domain-specific modifications were made to the harness. It solved the problem. It also improved on the human-written solution — tightening a key mathematical bound by more than fourfold and achieving complete coverage of the vertex set.
A system built to write code produced better mathematics than the mathematicians who set the problem. That is worth sitting with.
The finding isn’t that the harness is secretly a math genius. The finding is that the organizational architecture — decompose the problem, parallelize exploration, verify outputs, iterate toward completion — transfers across domains at near-zero cost.
Four teams, one pattern
What makes this more than an interesting anecdote is that four organizations arrived at the same structural pattern independently:
Cursor uses a Planner-Worker-Judge hierarchy: planners decompose and assign, workers grind individual sub-tasks in isolation, a judge evaluates partial output and decides whether to continue or restart cleanly. Flat coordination — all agents at the same level — failed badly; agents became risk-averse and avoided hard problems.
Anthropic’s internal pattern separates an Initializer (sets up the environment and writes a progress file) from a Coder (makes incremental progress and leaves structured artifacts for the next session). Without it, the agent tries to one-shot the whole implementation, runs out of context, and reports done without testing.
Google DeepMind’s Aletheia system uses Generator-Verifier-Reviser: the agent that generates output and the agent that checks it have different attentional profiles, which is why separating them improves results.
OpenAI’s Codex runs parallel sandboxed environments with verification loops, sustaining up to seven-plus hours of autonomous work.
None of these teams coordinated. They converged because the pattern is latent in the structure of hard work itself. That’s the most reassuring sign that it’s real.
What this means in practice
The smoothing of the frontier is accelerating faster than model improvement alone would predict, because architectural insights transfer at near-zero cost and don’t depreciate. Cursor didn’t spend six weeks building a math harness. They spent a few days pointing an existing one at a math problem.
The practical question this raises for each of us: which parts of our work are actually verifiable — either by a machine, or by an expert applying clear criteria — and which parts genuinely require judgment that can’t be checked? Most people, when they map this out carefully, find far more in the “checkable with clear criteria” category than they expected. That’s where structured delegation becomes available almost immediately.
The jagged frontier was never AI’s ceiling. It was a measurement artifact from a period when we hadn’t yet learned how to ask.
We’re learning now, and the pace is picking up.
Text summarized and optimized using Anthropic’s models and reviewed by a human.