A team of AI coding agents run by a supervisor that keeps rebuilding the team wrote a GPU kernel that reached a 1.21x mean speedup over FlashInfer’s production MLA decode kernel, according to a UC Berkeley preprint posted to arXiv on 27 October. The authors say it was “faster on 36 of 47 workloads” from FlashInfer-Bench.
Under a matched token budget, the paper says, “Claude Code agent plateaus at 0.51x the expert-written FlashInfer kernel, while our system reaches 0.74x”. In its ablation, a fixed four-role team reached 0.31x. By our arithmetic, letting the supervisor restructure the team gave 2.4 times the fixed team’s result: 0.74 divided by 0.31.
Business Pill · AI Agents Explained
An evergreen Business Engineer explainer on AI agents. It teaches the general idea only and does not describe the paper’s methods.
The key insight: A UC Berkeley preprint reports that a supervisor that keeps rebuilding a team of Claude Code agents reached 0.74x FlashInfer’s MLA decode speed on a matched budget, against 0.51x for Claude Code alone and 0.31x for a fixed four-role team; by our arithmetic that is 2.4 times the fixed team.
The system, called Evolving Agent Teams, is by Shiyi Cao, Ziming Mao, Dacheng Li, Joseph E. Gonzalez and Ion Stoica of the University of California, Berkeley. The paper says it has “a central supervisor that diagnoses each local optimum and restructures a team of agents” and “a persistent file system that comprises a learning document, an evolution history, and per-team artifacts”.
It is built on Claude Code agents powered by Claude Opus 4.6 with 1M context, and the supervisor is itself a Claude Code agent. The paper says “There are no predefined initial roles or team configurations”; the supervisor designs all roles from the task description.
The problem it targets, per the paper: coding agents “often make rapid initial progress, but then converge to local optima far below expert-written baselines”. On MLA decode, Claude Code “abandons the difficult WGMMA tensor core implementation after repeated failures at correctness”.
Speedup on the MLA paged decode kernel (H100) against FlashInfer’s production kernel, from Sections 1 and 4.5 of arXiv:2610.36848 (27 Oct 2026). The 1.21x run used additional budget beyond the matched comparison. The 2.4x ratio (0.74 / 0.31) is our arithmetic.
What Was Posted on X
New UC Berkeley paper shows that a supervisor that keeps reorganizing a team of coding agents can beat expert GPU kernels where single agents stall.
Coding agents make fast early progress on GPU kernels, then settle on easier designs and plateau far below expert code. On the MLA decode kernel, Claude Code stalled at 0.51 times FlashInfer’s speed.
Evolving Agent Teams adds a supervisor that diagnoses each plateau and rebuilds the team, inventing roles like a profiler, a debugger, or a knowledge integrator. A shared file system keeps lessons from every past team.
Under the same token budget it reached 0.74 times FlashInfer on MLA decode, against 0.31 for a fixed 4-role team. With more budget it hit 1.21 times FlashInfer’s speed, winning on 36 of 47 production workloads.
For long, hard optimization runs, have a supervisor restructure agent roles when progress stalls instead of running 1 fixed agent setup.
Rohan Paul highlighted the paper on X at 21:22 UTC on 9 November, summarising that Claude Code alone stalled at 0.51 times FlashInfer’s speed on the MLA decode kernel, that Evolving Agent Teams reached 0.74 times against 0.31 for a fixed 4-role team under the same token budget, and 1.21 times with more budget. The advice to have a supervisor restructure agent roles when progress stalls is the post’s own.
What the Supervisor Changed
On the MLA decode kernel the system “explores over 100 team configurations”, the paper says. Its case study lists roles the supervisor added: a Disciplinarian that “blocks non-WGMMA commits”, a focused debugger running in parallel with an optimizer, and a Knowledge Integrator.
The Knowledge Integrator came after a recommendation was lost between teams. The paper’s case study says: “Lost knowledge between teams is the biggest structural weakness”. The supervisor then mandated that all future teams read a shared synthesis document.
On the ablation, the paper says the fixed team and the evolving team “both use the same model and multi-agent structure, but only the evolving system can restructure roles and coordination in response to diagnosed failure modes”. Codex with GPT-5.5 xhigh reached 0.28x.
Who Flagged It
Rohan Paul highlighted the paper in a post on X on 9 November, advising that for long, hard optimization runs teams should “have a supervisor restructure agent roles when progress stalls instead of running 1 fixed agent setup”.
As we read the paper, the gain it isolates sits in the organisation of the agents rather than the model: the ablation compares two multi-agent setups, and only the one whose supervisor could restructure roles moved past 0.31x.
As we read the case study, the supervisor’s fixes were management fixes: enforcing a rule on commits, splitting debugging from optimisation, and making every new team read a shared synthesis document.
Cao, Mao, Li, Gonzalez and Stoica, arXiv:2610.36848, 27 October 2026
“Lost knowledge between teams is the biggest structural weakness”
Three Implications
FOR TEAMS RUNNING CODING AGENTS The paper reports that a single Claude Code agent plateaued at 0.51x on MLA decode after abandoning the harder WGMMA design.
FOR AGENT-TOOL BUILDERS In the paper’s ablation, a fixed four-role team reached 0.31x and the evolving team 0.74x; the paper attributes the gap to restructuring roles and coordination.
FOR COMPUTE BUYERS With additional budget the team’s kernel reached 1.21x the production FlashInfer kernel, faster on 36 of 47 workloads, per the paper.
The framework says: “The agentic surface — the harness — is becoming where value pools in the entire AI stack.”
As we read it, this paper puts a number on that layer: the same family of Claude Code agents scored 0.31x in a fixed team and 0.74x when a supervisor could reorganise them. The framework predates the paper and does not discuss it.
What Is Not Established
This is a preprint and has not been peer reviewed. The authors say they run each main setting twice, which “limits statistical comparisons”. The 1.21x result used additional budget beyond the matched-budget comparison, so the 0.51x, 0.74x and 0.31x figures and the 1.21x figure come from different budgets. The paper names the fixed team’s model as Claude Opus 4.7 in its setup and Opus 4.6 for the evolving team. The 2.4 times figure is our arithmetic.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
A UC Berkeley preprint reports that a supervisor-led team of Claude Code agents reached 0.74x FlashInfer’s MLA decode speed on a matched budget, against 0.51x for Claude Code alone and 0.31x for a fixed team, and 1.21x with more budget, according to the paper.
A note on sourcing. Findings, figures and quotations are from the arXiv preprint 2610.36848 (27 October 2026) by UC Berkeley researchers, which has not been peer reviewed. The 2.4 times ratio is our arithmetic from the paper’s figures. The advice to have a supervisor restructure agent roles is Rohan Paul’s, from a post on X.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
