Week in review · Mon 21 Sept – Sun 27 Sept · 157 stories read

Agents, Benchmarks, and Embedded Hacks

This week the AI world kept building smarter agents and tougher benchmarks, while the hardware side saw a flood of open‑source tools and security patches. Engineers got fresh data to train on and new boards to tinker with.

By the numbers159.3Mparameters in the Mizar audio‑language model By the numbers52.92%mean accuracy of Mizar on the MM benchmark By the numbers6.8 GBsize of the archive Meta’s Muse AI sent back

Agent Harnesses Get a Makeover

Lead‑in: Researchers are converging on the idea of a reusable "harness" for LLM agents. Papers like Grow the Harness, Harness‑Zero, and RRSI propose distilling, regularizing, and iteratively improving shared control logic, while MAGIC and COMPASS explore graph‑based coordination and spatial transformers for multi‑agent teamwork. Together they signal a shift from monolithic prompts to modular, reusable agent components.

Grow the Harness, Not the Context: From Strategy-Free Scaffolds to Reusable Specialist AgentsWed 23 Sept · arXiv cs.AI
Harness-Zero: Harness Distillation via Agent-as-HarnessTue 22 Sept · arXiv cs.AI
RRSI: Regularized Recursive Self-Improvement of Agent HarnessesTue 22 Sept · arXiv cs.LG
MAGIC: Mixed-Granularity Agent Graphs via Incremental Construction with Dense-Reward Reinforcement LearningWed 23 Sept · arXiv cs.LG
Controlling Collectives of AI Agents in Reasoning Space with Spatial TransformersThu 24 Sept · arXiv cs.RO

Benchmarks Go Domain‑Specific

Lead‑in: New evaluation suites are targeting narrow, high‑impact domains. BRIE offers a living benchmark for EHR question answering, validated by clinicians, while SWE‑Serve provides 53 production‑inference tasks for agentic engineering. These datasets push models to prove usefulness beyond generic NLP, emphasizing reliability in health and deployment contexts.

A Living Benchmark for Information Retrieval from Electronic Health RecordsFri 25 Sept · arXiv cs.AI
SWE-Serve: Benchmarking Agentic Engineering For Production Inference ServingWed 23 Sept · arXiv cs.AI

Embedded DIY and Security Surge

Lead‑in: The hardware community is buzzing with open‑source releases and critical patches. CERN’s colibri VHDL library, ESP32‑based audio spectrum and lighting projects, and a modular macro keypad showcase the maker spirit, while Friday’s multi‑distro security updates and the Linux 7.3‑rc4 prepatch remind engineers to stay patched.

FPGA for All — CERN Releases “colibri” VHDL LibraryFri 25 Sept · hackaday
Audio Spectrum Analyzer on an ESP32 Display BoardThu 24 Sept · hackaday
ESP32 Replacement for Lighting Display ControllerWed 23 Sept · hackaday
A Modular Macro KeypadSun 27 Sept · hackaday
Security updates for FridaySat 26 Sept · lwn

Systems Research Tackles Memory and Consensus

Lead‑in: On the CS front, xTier demonstrates kernel‑resident learned memory tiering for CXL servers using eBPF, while Hydrozoan and Flamingo propose latency‑adaptive DAG consensus and load‑balancing for Byzantine fault‑tolerant systems. These works aim to make large‑scale infrastructure both faster and more resilient.

xTier: Intelligent Tiering for CXL-Enabled MemoryThu 24 Sept · arXiv cs.OS
Hydrozoan: Latency-Adaptive DAG Consensus under Mixed Byzantine and Crash FaultsWed 23 Sept · arXiv cs.DC
Flamingo: On Load Balancing in DAG-based Consensus ProtocolsThu 24 Sept · arXiv cs.DC

Watch next weekNext week keep an eye on how the new agent harness frameworks are adopted in production benchmarks and whether the emerging CXL tiering tricks make it into mainstream kernels.

Today’s edition 2026-W40 →