Summary
No curated reposts or saved posts landed in the morning slot, so this synthesis comes from the morning capture of the Following feed, which covers the US Friday evening and Saturday. The strongest signal is a cluster of five posts on agent harnesses and who makes each decision inside them. DAIR.AI's pick of JevAgentHarness is the anchor: a 4B decision model takes an agent's routine tool and argument choices when it is confident, and the big model handles the rest, with 14.2% lower cost and no changed outcomes across 500 tasks. Google's open-sourced Harness², a 14-lecture harness-engineering course and a Harvard/MIT paper on agents that bid for each step fill out the cluster. A second, smaller cluster is agent training: NVIDIA's midtraining recipe for coding agents, CoPE's one-loss fix for GRPO sample efficiency and Salesforce's diverse-ideas model. The standout hardware item is a claim that B300 rental prices rose more than 20% in a month even as neocloud stocks fell. Hype threads about "reverse-engineer any app" tools made up a large share of the high-engagement posts and are skipped.
Posts
Decision models inside the harness (cluster of 5) (@dair_ai, @omarsar0, @0x_rody, @dr_cintas, @techNmak; paper). JevAgentHarness (UIUC) sends each multiple-choice decision an agent makes, such as which tool to call or which argument to pass, to a fast decision model. It uses the answer only when confidence clears a threshold fitted on that agent's own past runs. With Gemini 3.5 Flash as the agent, the decision model made about a quarter of tool choices and cut cost 14.2% without changing any of 500 task outcomes. A fixed 0.5 threshold lost 21.5 points. Its open AgentDecision-4B, trained on states the agent reaches after its own mistakes, beat its teacher (91.5% vs 82.0% on shifted tasks) and made per-agent runs 81-89% cheaper. A practitioner post claims wiring Jev into Grok Bot cut his "fork" spend from about $110 to $4 a day (unverified). Google Research open-sourced Harness², a framework that refines the harness around a fixed model at both task and meta level, built from eight parts (system prompt, guardrails, memory, tool guidance, subagents, skills, scripts, plugins). Walking Labs published a free course on harness engineering. See JevAgentHarness summary.
Agents that bid for each step (@snwiki238337; paper). A long Chinese-language thread on Harvard and MIT's market-based post-training (MPT). Each agent learns an action policy and a bidding policy, the highest bidder acts, and training credit is split by bid, so an agent that bids high and fails is penalized most. Trained swarms improve as more agents join at inference and degrade less when agents drop out. See MPT summary.
Agent training recipes (cluster of 3) (@omarsar0, @haofeiyu44, @dair_ai). Salesforce trains an 8B model to write a set of research ideas in one pass, scoring each idea by quality times distance from earlier ones. It beats Claude Opus 4.8 and GPT-6 Astra on "creative utility" (paper). CoPE adds one loss to GRPO: on successful runs, the policy also predicts its next few actions, which gives up to 2x sample efficiency with no inference cost (paper). NVIDIA's midtraining stage of 500B software-engineering tokens at 256K context lifts SWE-bench Multilingual from 55.8 to 62.8 on Nemotron 3.5 Lightning (paper). See recipes summary.
Hy-Memory and research-agent forgetting, recirculating (@dair_ai, @omarsar0). Tencent's 3B-active memory writer, which matched Claude Opus 5 inside four memory frameworks, and Berkeley's finding that research agents scale further when they periodically discard their history, both covered earlier this week, kept circulating. See the 10-17 digest.
B300 rentals up, neocloud stocks down (@StockSavvyShay). An investor account says B300 rental rates rose more than 20% over the past month while Nebius, IREN and CoreWeave sold off on fears that GPUs lose value too fast. It credits demand for open-model inference. The figure is unsourced.
Next-Concept Modeling (@TheTuringPost; paper, code). Stanford and Together's third way to generate text, coarse-to-fine through a hierarchy of compressed codes, resurfaced with its code release. Covered on 10-16.
Product notes (@cohere, @alexandr_wang, @github). Cohere says Embed 5 Fast beats Google's EmbeddingGemma 2 on over 95% of 163 datasets by its own benchmark. Alexandr Wang restated Muse Spark Open's specs (816B total, 42B active, 1M context). GitHub paired its Copilot app with a Logitech keypad for mapped agent prompts.
Software-engineering automation, two views (@fchollet, @AnatoliKopadze). Chollet notes there are more software engineers than in 2025 and they work harder, even though they mostly no longer write code. A clip of the head of Claude Code says he builds graphs and loops that build agents, rather than prompting them.
Skip. The "REA reverse-engineers any app" and "Compaq cloned IBM with one prompt" threads, a "100 free APIs" list, vibe-coder security lists, Forward Deployed Engineering career advice, an unverified "OpenAI 10-page always-on agents PDF" thread, and off-topic posts.