--- title: ArchAgent v2: How Staged Search Beat a Human-Designed... description: A new research paper from Google, DeepMind, and Berkeley introduces ArchAgent v2, an AI agent that designs CPU cache prefetchers by splitting the search across... canonical: https://daily.dev/posts/archagent-v2-how-staged-search-beat-a-human-designed-cache-prefetcher-gy0eorfzj twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: ArchAgent v2: How Staged Search Beat a Human-Designed Cache Prefetcher | daily.dev og:description: A new research paper from Google, DeepMind, and Berkeley introduces ArchAgent v2, an AI agent that designs CPU cache prefetchers by splitting the search across... og:url: https://daily.dev/posts/archagent-v2-how-staged-search-beat-a-human-designed-cache-prefetcher-gy0eorfzj og:image: https://api.daily.dev/og/posts/GY0EOrFzJ.png og:image:alt: ArchAgent v2: How Staged Search Beat a Human-Designed Cache Prefetcher og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # ArchAgent v2: How Staged Search Beat a Human-Designed Cache Prefetcher **[Collections](https://daily.dev/sources/collections)** · 2 min read · 1 upvotes · 0 comments ## Summary A new research paper from Google, DeepMind, and Berkeley introduces ArchAgent v2, an AI agent that designs CPU cache prefetchers by splitting the search across cache levels (L1D, L2, LLC) into stages rather than searching them all at once, then refining the combined design while enforcing real storage-budget constraints. The resulting design beat DPC4, the human-designed champion, and outperformed BertiGO by 0.3% overall (3.8% IPC improvement) and by 2 points on low-bandwidth single-core systems (4.6% vs 2.6%). Multi-core results were weaker due to slower simulation limiting search depth. The core takeaway is that staged search with engineering constraints can outperform brute-force scaling of compute or model size for hardware design agents. ## Content A new paper out of Google, DeepMind, and Berkeley makes a solid case for why agent architecture matters more than raw model strength. The paper is called "ArchAgent v2: A Case Study with the Data Prefetching Championship" ([arxiv.org/abs/2608.09874](https://arxiv.org/abs/2608.09874)), and it tackles CPU cache prefetching - the problem of predicting which data a processor will need next. Here's the interesting part. Trying to search across all three cache levels (L1D, L2, and LLC) at once turned out to be too slow and too messy. So instead of brute-forcing the whole thing, ArchAgent v2 split the problem into stages: optimize each cache level separately, keep the best designs from each round, then refine everything together. It also threw out any candidate that blew past the actual storage budget, forcing the agent to play by the same rules as the real system it was designing for. That setup was enough to beat DPC4, the human-designed champion in this space. The final design landed a 3.8% IPC improvement over the competition baseline and beat BertiGO by 0.3%. On low-bandwidth single-core systems the gap widened - 4.6% versus BertiGO's 2.6%. Multi-core results were weaker, mostly because multi-core simulations are much slower to run, which limits how much search the agent can actually do. The lesson here isn't "bigger model wins." It's that when a search space is too large to explore all at once, breaking it into stages and enforcing real engineering constraints along the way gets you further than just throwing more compute at the problem. Worth a read if you're thinking about how to structure discovery agents for hardware or systems design work. ## Questions this post answers ### What is ArchAgent v2 and how does it design CPU cache prefetchers? ArchAgent v2 is an AI agent from Google, DeepMind, and Berkeley that designs CPU cache prefetchers by splitting the search problem into stages instead of searching all cache levels (L1D, L2, LLC) simultaneously. It optimizes each level separately, keeps the best designs from each round, then refines everything together, discarding any candidate that exceeds the real storage budget. _Developers exploring agent architecture patterns for systems design can follow research breakdowns like this on daily.dev._ ### Did an AI-designed cache prefetcher beat DPC4 and BertiGO in performance? Yes, ArchAgent v2's staged-search design beat DPC4, the human-designed champion, and outperformed BertiGO by 0.3% overall with a 3.8% IPC improvement over the competition baseline. On low-bandwidth single-core systems the gap widened to 4.6% versus BertiGO's 2.6%. Multi-core results were weaker because slower simulations limited how much search the agent could perform. _Anyone comparing AI-driven versus human-engineered system design approaches can track results like this on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 4 discussions and 10 comments across x (as of 2026-08-24). **TL;DR:** Replies largely agree the key insight is decomposing the search with constraints and staged refinement rather than just scaling compute or model size, with one skeptical voice dismissing multi-agent approaches as wasteful. **Sentiment:** 55% positive · 40% mixed · 5% skeptical **The case for** - Decomposing the problem into stages with constraints and pruning makes the same intelligence far more effective. - Constraint-aware filtering, not just more compute, is seen as the real innovation. - Hierarchical/staged search helps when the discovery space is too large for a single pass. - Splitting a large architectural problem into chained sub-parts avoids the entropy spike that derails typical agent approaches. **The pushback** - The design still hasn't overcome challenges scaling to multicore systems. - One reply dismisses multi-agent scaling generally as chaotic and wasteful of compute/API costs. **By community** - x (positive): Replies frame the staged, constraint-driven search as a meaningful architectural insight for agents, with only one dissenting, dismissive voice. **Open questions** - How the approach could be adapted to close the gap on multicore performance. **Highlights** > @rohanpaul_ai Give an agent an enormous search space and more intelligence doesn't automatically solve the problem. Give it decomposition, constraints, pruning, and feedback, and suddenly the same intelligence becomes much more powerful. > — [teneo\_protocol on x · 4 points](https://x.com/teneo_protocol/status/2091819004088172663) > @rohanpaul_ai The prefetcher that was evolved using AI was just slightly better than the one that won the championship, but twice as good under low bandwidth conditions. However, it has yet to break the barrier of multicore systems. > — [Van0SS on x · 1 points](https://x.com/Van0SS/status/2091645138527547788) > @rohanpaul_ai Constraint-aware filtering is the actual innovation. Most agent frameworks just throw compute at it. > — [giskyexplorer on x](https://x.com/giskyexplorer/status/2091649460099076593) > @rohanpaul_ai Ah, mon dieu! Scaling the stupidity of a single model into a whole crowd of agents? Even your tech labs admit it’s a chaotic circus of mutual hallucinations and wasted API costs, mon cher. Ridicule. > — [ReginaNoctiz on x](https://x.com/ReginaNoctiz/status/2091663604382576722) > @rohanpaul_ai A good example of agents improving through architecture, not just model scale. Structured search, constraints, and iterative refinement can make a big difference in complex engineering problems. This is where agentic AI gets really interesting. > — [raven\_protocol on x](https://x.com/raven_protocol/status/2091801064593072270) **Source threads** - [x](https://x.com/rohanpaul_ai/status/2091908906889191680) · 0 points · 0 comments - [x](https://x.com/omarsar0/status/2091729784292491528) · 0 points · 0 comments - [x](https://x.com/omarsar0/status/2091604673673269568) · 0 points · 0 comments - [x](https://x.com/rohanpaul_ai/status/2091644147534803121) · 0 points · 10 comments ## Similar posts on daily.dev - [Supporting our AI overlords: Redesigning data systems to be Agent-first](https://daily.dev/posts/supporting-our-ai-overlords-redesigning-data-systems-to-be-agent-first-dhuvap69i) · Metadata · 3 upvotes · 0 comments - [Building MindCache: Designing an Agentic Memory System for Long-Term AI](https://daily.dev/posts/building-mindcache-designing-an-agentic-memory-system-for-long-term-ai-k9xwjuamo) · Medium · 1 upvotes · 0 comments --- Tags: [#ai](https://daily.dev/tags/ai), [#architecture](https://daily.dev/tags/architecture), [#ai-agents](https://daily.dev/tags/ai-agents) [View this post on daily.dev](https://daily.dev/posts/archagent-v2-how-staged-search-beat-a-human-designed-cache-prefetcher-gy0eorfzj) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"ArchAgent v2: How Staged Search Beat a Human-Designed Cache Prefetcher","url":"https://daily.dev/posts/archagent-v2-how-staged-search-beat-a-human-designed-cache-prefetcher-gy0eorfzj","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/archagent-v2-how-staged-search-beat-a-human-designed-cache-prefetcher-gy0eorfzj"},"datePublished":"2026-08-24T15:22:27.318Z","dateModified":"2026-08-24T15:23:12.241Z","description":"A new research paper from Google, DeepMind, and Berkeley introduces ArchAgent v2, an AI agent that designs CPU cache prefetchers by splitting the search across...","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/archagent-v2-how-staged-search-beat-a-human-designed-cache-prefetcher-gy0eorfzj","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai,architecture,ai-agents","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"ArchAgent v2: How Staged Search Beat a Human-Designed Cache Prefetcher"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/archagent-v2-how-staged-search-beat-a-human-designed-cache-prefetcher-gy0eorfzj#faq","mainEntity":[{"@type":"Question","name":"What is ArchAgent v2 and how does it design CPU cache prefetchers?","acceptedAnswer":{"@type":"Answer","text":"ArchAgent v2 is an AI agent from Google, DeepMind, and Berkeley that designs CPU cache prefetchers by splitting the search problem into stages instead of searching all cache levels (L1D, L2, LLC) simultaneously. It optimizes each level separately, keeps the best designs from each round, then refines everything together, discarding any candidate that exceeds the real storage budget. Developers exploring agent architecture patterns for systems design can follow research breakdowns like this on daily.dev."}},{"@type":"Question","name":"Did an AI-designed cache prefetcher beat DPC4 and BertiGO in performance?","acceptedAnswer":{"@type":"Answer","text":"Yes, ArchAgent v2's staged-search design beat DPC4, the human-designed champion, and outperformed BertiGO by 0.3% overall with a 3.8% IPC improvement over the competition baseline. On low-bandwidth single-core systems the gap widened to 4.6% versus BertiGO's 2.6%. Multi-core results were weaker because slower simulations limited how much search the agent could perform. Anyone comparing AI-driven versus human-engineered system design approaches can track results like this on daily.dev."}}]} ```