Evolving Agent Teams
Comments on “Evolving Agent Teams”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Berkeley researchers describe Evolving Agent Teams, a framework in which a supervisor agent diagnoses when coding agents plateau and restructures roles, coordination and shared notes in a persistent file system. On FlashInfer's MLA paged decode kernel, it reaches 0.74x of the expert kernel at a matched budget, against 0.51x for Claude Code, and 1.2x with more budget, winning 36 of 47 workloads. Results are the best of two runs, and the claim check found no independent sources for the figures, so they rest on the authors' own reporting.
Why it ranked
No. 8 of 59 with a score of 0.66. The panel rated it highest on new information (0.92) and technical depth (0.85), lowest on audience breadth (0.59). Placed No. 8, No. 12 by score alone: the placement terms moved other stories past it.
- AI judgeJudge 2claude-haiku-5-5
- Academic paper on multi-agent LLM teams evolving to optimize GPU kernels; substantive benchmarks with self-reported results, a novel framework, mostly relevant to AI and performance engineering, with some overclaim risk and no prior coverage.
- AI judgeJudge 3gpt-5.6-luna
- A detailed research paper presents a novel evolving multi-agent framework, extensive GPU-kernel experiments, comparisons, and reproducible technical mechanisms. It is highly relevant to AI and systems builders, though limited runs and broad cross-domain claims temper certainty.
Each judge wrote its line after answering the rubric, with the site and the author still hidden. Every answer behind this is in the table under Score.
Score
0.66total under rubric v2.3
- Significance0.69
- Reader interest0.78
- Audience breadth0.59
- Practical relevance0.66
- Substantiated0.78
- Technical depth0.85
- New information0.92
- Counts against: Promotional0.07
- Counts against: Overclaim0.39
Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).
| Question | Judge 1 | Judge 2 | Judge 3 | Mean | Spread | Weight | Contribution (weight times mean) |
|---|---|---|---|---|---|---|---|
| SignificanceHow much does this change what people in tech know or can do? | 0.70 | 0.55 | 0.82 | 0.69 | 0.27 | +0.24 | +0.166 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.71 | 0.70 | 0.93 | 0.78 | 0.23 | +0.26 | +0.203 |
| Audience breadthHow many people who build technology does it concern? | 0.65 | 0.50 | 0.62 | 0.59 | 0.15 | +0.13 | +0.077 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.66 | 0.55 | 0.78 | 0.66 | 0.23 | +0.22 | +0.146 |
| SubstantiatedHow well is it backed? Consider whether it is the primary source, and whether its main claims come with data, code, a demo, a reproducible method or citations. | 0.96 | 0.60 | 0.78 | 0.78 | 0.36 | +0.05 | +0.039 |
| Technical depthHow technically substantive is it? | 0.90 | 0.70 | 0.96 | 0.85 | 0.26 | +0.05 | +0.043 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.91 | 0.90 | 0.94 | 0.92 | 0.04 | +0.05 | +0.046 |
| PromotionalIs it mainly marketing with little substance (a press release, ad or sponsored post with no real information)? A launch post that carries real news is not promotional. | 0.08 | 0.10 | 0.04 | 0.07 | 0.06 | −0.19 | −0.014 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.49 | 0.35 | 0.32 | 0.39 | 0.17 | −0.11 | −0.043 |
| Sum of contributions | +0.662 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.662 | ||||||
Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.
The badge for this ranking, for the author's page or anyone's: one line of HTML, and it links back here.
<a href="https://rubric.hrih.si/editions/2026-11-01/8/"><img src="https://rubric.hrih.si/badge/2026-11-01/8.svg" alt="Ranked #8 on Rubric, 1 Nov 2026" width="231" height="28"></a>Claim check
Claim check: 7 no source given · 1 couldn’t check. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.
No source givenBenchmark result
With an extended token budget, Evolving Agent Teams generates a CUDA kernel reaching 1.2x mean speedup over the production FlashInfer MLA paged decode kernel and wins on 36 of 47 production workloads from FlashInfer-Bench.
The article gives no link or citation for it.
The article says “generates a CUDA kernel that reaches $1.2\times$ mean speedup over the production FlashInfer MLA paged decode kernel from scratch and wins on 36 of 47 production workloads”
The source says No source to compare with.
No source givenBenchmark result
On MLA paged decode, the Claude Code agent plateaus at 0.51x the expert-written FlashInfer kernel while Evolving Agent Teams reaches 0.74x under a matched token budget.
The article gives no link or citation for it.
The article says “Claude Code agent plateaus at $0.51\times$ the expert-written FlashInfer kernel, while our system reaches $0.74\times$ under a matched token budget”
The source says No source to compare with.
No source givenBenchmark result
The previous best result on the MLA decode task is 0.61x from K-Search.
The article gives no link or citation for it.
The article says “while the previous best result is $0.61\times$ from K-Search”
The source says No source to compare with.
No source givenNumber
Codex 5.5 xHigh reaches only 28% of expert performance on the MLA decode kernel despite a $1K budget and multiple days of effort.
The article gives no link or citation for it.
The article says “Codex 5.5 xHigh reaches only 28% of expert performance despite a $1K budget and multiple days of effort”
The source says No source to compare with.
No source givenBenchmark result
On KDA prefill, the team's kernel is 3.01x faster than vLLM's hand-written HIP kernels on AMD MI350X and 3.44x faster than FlashKDA on B200.
The article gives no link or citation for it.
The article says “the team’s kernel is $3.01\times$ faster than vLLM’s hand-written production HIP kernels [17] on AMD MI350X, and $3.44\times$ faster than FlashKDA”
The source says No source to compare with.
No source givenBenchmark result
On RoboDojo, the team's code policy scores 62.1 on held-out seeds, above the best leaderboard entry of 55.3 as of Oct. 26, 2026.
The article gives no link or citation for it.
The article says “scores $62.1$ on held-out seeds, above the best leaderboard entry ( $55.3$ as of Oct. 26, 2026”
The source says No source to compare with.
No source givenCausal conclusion
The fixed four-role team reaches 0.31x on MLA decode, versus 0.74x for the evolving team; the authors say this gap isolates the value of team evolution.
The article gives no link or citation for it.
The article says “The $0.31\times\to 0.74\times$ gap between the fixed team and the evolving team isolates the value of team evolution”
The source says No source to compare with.
Couldn’t checkBenchmark result
On B300, the team's final kernel scores 3.54x in the same harness, against 3.02x for the agent-written KDA2 kernels.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “On B300, the final kernel scores $3.54\times$ in the same harness, against $3.02\times$ for the agent-written KDA2 kernels”
The source says No passage could be compared.
Source cited by the article: nvlabs.github.io/kda/blog/2026-09-27-kda-for-kda
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN 1 pts, not on front page
Recorded for context only, checked 1 Nov 2026, 05:47 UTC. It played no part in the ranking.