Items in this briefing come from Hugging Face Daily Papers, MIT News, Simon Willison and arXiv.org.
The Signal
Getting individual decisions right may not be enough for an agent to finish its tasks. The tool memory authors report that in a test on three models, many failures persisted even when the correct next tool was supplied. The authors of a paper on small decision models report that the model most accurate on isolated questions still failed some agent tasks.
Via: Simon Willison (blog post) · Published: 2026-10-15
Why this matters: Readers following Anthropic's finances can hold these figures from Anthropic's investor event, which come from unnamed sources in the Financial Times report, up against anything the company later publishes.
Summary: This is a blog post by Simon Willison summarizing a Financial Times article about Anthropic's investor event on Wednesday. According to that article, which cites people with knowledge of the matter, Anthropic told investors its revenue was about $4.8bn in the first quarter, $11.5bn in the second and $16.8bn in the third. In the third quarter it reportedly earned $1.70 in revenue for every $1 spent on computing power. Adjusted operating margin, excluding stock-based compensation, was reportedly 11%, though the post gives no period for it. Willison lists these as key things he learned from the article.
Why this matters: Teams building agents that use tools and keep graph-style memory can check whether any gains come from the memory or from the agent simply making more tool calls. One way to check is to test a plain instruction.
Summary: This arXiv preprint tests "tool graph memory," a method where AI agents store past tool-to-tool transitions to guide which tool to call next. The authors ran nine open-weight models on two multi-turn benchmarks and report that two such methods seldom beat a memory-free baseline by a statistically significant amount. One of them used about 2.7 times as many model calls on one benchmark. Random or removed suggestions kept most of its gain, and a simple instruction to keep trying tools did about as well. Storing the same experience as written insights did better with almost no extra calls. The authors report that many agent failures persisted even when the correct next tool was supplied, such as giving up without calling a tool. They note that their deeper analyses, this test included, cover only three models on one benchmark.
Via: Hugging Face Daily Papers (paper) · Published: 2026-10-14
Authors: Kunlun Zhu et al.
Why this matters: Developers trying to cut the cost of running AI agents can see from these results that handing tool choices to a small model needs a tuned confidence threshold to avoid losing task success.
Summary: A research paper looks at cutting the cost of AI agents by letting a small decision model answer multiple-choice steps, such as which tool to call next, instead of a large model. The authors built a test suite and report that the decision model most accurate on isolated questions solved only 69.5% of agent tasks. In a separate result, delegating tool choices without a fitted confidence threshold lowered success by 19.0 points. They then built a system that fits that threshold from the agent's own runs, and trained a small open model on states agents reach after mistakes. On tasks unlike its training data, with the decision model making every tool choice, they report agents solved 91.5% with their model and 82.0% with an existing decision model. The large model alone solved all of them. On tasks like the training data, they report agents were 81.1 to 89.0% cheaper and up to 2.5 times faster, without losing 3 or more points of task success.
Why this matters: Researchers who grow stem cells or lab-made organ models, including lung models the article says would allow studies of asthma, lung cancer and pulmonary fibrosis, may be able to use these signal patterns to guess which signals to test, rather than mapping every pathway in every cell type.
Summary: This MIT News report describes a study from a team led by Pulin Li and Nicholas Hutchins, published in Nature Methods on Sept. 8. They report that each cell signaling pathway leaves a distinctive pattern of gene activity, and that this pattern is consistent across cell types. A neural network model called IRIS was trained on data from human embryonic stem cells exposed to many combinations of six signaling pathways. It was then tested on single cells from mouse embryos and, according to the report, accurately predicted when and where pathways switched on in cells headed for heart, gut, muscle and spinal cord tissue. The model also predicted that activating one pathway would promote lung-specific development, and experiments in mouse embryos confirmed this.
Sources:
Hugging Face Daily Papers, MIT News, Simon Willison, arXiv.org
Daily Signals currently covers one topic: Artificial Intelligence. More topics may follow.