4 AI Labs Built the Same System Without Talking to Each Other (And Nobody's Discussing Why)

AI’s “jagged” performance pattern—great at some tasks, weak at others—is increasingly an artifact of how systems are deployed, not a permanent feature of intelligence. As agentic workflows mature, work that can be decomposed, parallelized, verified, and iterated is getting smoother, meaning fewer abrupt failures and less whiplash between “it works” and “it doesn’t.” That shift matters because it changes what organizations should expect AI to handle in day-to-day roles like PRDs, codebases, and customer service—areas where iterative human-style processes already exist.

The core diagnosis is about single-turn, single-agent prompting. When a model must answer in one shot, errors propagate and there’s no built-in mechanism to detect mistakes, retry, or accumulate information beyond a context window. That setup forces “oneshot cognition” onto problems that professionals normally solve through drafts, feedback loops, intermediate checkpoints, and revision cycles. For years, AI was placed into that primitive interaction style, and the resulting uneven outcomes were treated as evidence of uneven intelligence.

Over the last year, multiple improvements have started to smooth that curve. Inference-time compute lets models spend more time thinking, try alternative tokens, and correct course—an approach associated in the transcript with “ChatGPT 5.2 thinking” and “5.2 Pro.” But the bigger change is organizational: better tooling, better prompting, and—most importantly—agentic “harnesses” that provide scaffolding around an agent so it can do long-running work. A harness can include task files, memory, and the operational environment that lets an agent operate like a team rather than a lone responder.

The strongest proof offered is a March 3 announcement from Cursor CEO Michael Trule: Cursor reportedly found a novel solution to “problem six” of the first proof, improving on an official human-written solution with stronger bounds and better coverage. The system reportedly ran for four days with zero hints, zero human nudges, and zero midcourse guidance, using the same coding harness that earlier built a web browser from scratch. The significance isn’t just that it solved a math problem; it did so using a system designed for coding, suggesting that the coordination-and-verification architecture generalizes beyond software.

Cursor’s earlier work on long-running autonomous coding points to the mechanism: flat coordination failed because agents became risk-averse and avoided difficult work. The breakthrough came from hierarchy and specialization—planners decompose tasks into subplans, workers execute in isolation, and a judge role decides whether to continue, restarting cleanly to avoid context-window collapse. The transcript also claims model choice matters for long-horizon tasks, with GPT 5.2 outperforming Claude Opus in this setup.

The broader claim is convergence across organizations: Anthropic, Google DeepMind, OpenAI, and Cursor independently built multi-agent coordination systems with similar structural patterns—decompose, parallelize, verify, and iterate. The “jaggedness” narrative, then, is replaced by a management narrative: organizational intelligence scales when roles, handoffs, and verification are built into the system. The practical takeaway is that AI will be most useful where answers are machine-checkable or expert-checkable, and where workers can shift from “doing” to “sniff-checking” correctness and maintainability. The transition is expensive in tokens and effort, but it’s positioned as the path to smoother AI performance for real workplace tasks.