If agents work with other agents, they need a standard way to talk. That was the case for an explicit agent-to-agent protocol. Adoption is real. Meanwhile the rest of the stack has reduced how often agents need to appear to one another as independently addressable peers.

A remote agent can look like a capability. Inference stays behind the interface. Subagents live inside one runtime and die with the task. A supervisor presents one identity while specialists stay internal. Software factories coordinate the work, not a mesh of persistent agent relationships. Workflow engines treat agents as steps in a graph they already own. Handoffs can be artifacts in durable state rather than messages. Queues and event buses already do this at scale.

Most agent frameworks split the system across prompt templates, tool schemas, callback code, and workflow graphs. The result is fragile and hard to reason about.

A simpler model is possible. Treat the agent as an ordinary Python object. Its methods are the actions the model can take. Its fields hold state. Its docstrings serve as prompts. Its type annotations act as contracts. Methods whose body is an ellipsis are completed at runtime by an LLM-driven loop. Methods with normal bodies remain ordinary deterministic Python.

A new generation of general-purpose "AI scientist" tools is emerging, capable of helping with literature reviews, data analysis, hypothesis generation, figure creation, and even manuscript drafting. These agentic systems go beyond narrow tools like AlphaFold by breaking down complex requests into steps and integrating external software.

Nature's Ewen Callaway offers a grounded guide for researchers navigating this rapidly evolving landscape. Key recommendations include:

Wearable devices generate enormous amounts of multimodal sensor data (heart rate, motion, skin conductance, temperature, etc.), but most AI models for health are narrow, task-specific, and require large amounts of labeled data. This limits scalability and generalization in real-world settings where data is often fragmented or incomplete.

Google researchers introduce SensorFM, a large sensor foundation model pre-trained self-supervised on over one trillion minutes of wearable data from more than five million consented participants across diverse devices and geographies. It learns a general-purpose representation of human physiology using a missingness-aware framework (AIM) that treats real-world data gaps as meaningful signals rather than noise.

Treatment reasoning in medicine is inherently iterative: it requires identifying missing information, gathering evidence from multiple sources, evaluating contraindications and comorbidities, and revising conclusions as new data emerges. Most current AI systems struggle with this because they lack a structured way to seek and integrate evidence before forming a recommendation.

Researchers introduce ATHENA-R1, an AI agent trained to perform treatment reasoning across all FDA-approved drugs since 1939 by interacting with a universe of 212 real biomedical tools (covering indications, interactions, adverse events, disease phenotypes, and more). The agent learns to identify knowledge gaps, call the appropriate tools, incorporate retrieved evidence, and iterate until it reaches a grounded decision.

AI models are getting extremely good at anything that can be turned into a clear loss function or well-defined, gradable task — schoolwork, LeetCode, standard engineering problems, and most routine coding.

In this environment, the highest-leverage human skills are shifting to the areas that are hardest to grade or automate within a single training loop:

The dominant conversation in AI still revolves around which company has the smartest frontier model. However, the real economic transition underway is from maximizing raw intelligence to maximizing intelligence per dollar.

Enterprises are rapidly discovering that most workloads do not require the most expensive frontier models. As inference costs continue collapsing at an extraordinary rate, organizations are shifting from asking "Does AI work?" to "Does it work economically at scale?" This change favors cheaper, "good enough" models (including open-weight ones) for the vast majority of tasks, while reserving frontier intelligence only for high-value problems where additional capability creates outsized returns.

Most LLM evaluations rely on generic benchmarks that often fail to predict performance in actual production agent systems. This experiment kept the entire agent infrastructure fixed (tools, prompts, retrieval, retry logic, and orchestration) and swapped only the underlying model across 15 LLMs.

The agent was tasked with turning natural language business questions into validated, executable queries against a large analytics store — a realistic, multi-step workflow involving planning, tool use, query generation, self-repair, and output validation.

Most current evaluations treat agent memory as a black box, measuring only end-to-end task success. This paper takes a data management perspective and decomposes agent memory into four core modules:

•⁠ ⁠Representation & Storage

Many optimization patterns from the mainframe era are being revived in agentic AI systems. When a core resource is powerful but expensive and slow, we wrap it in caching, abstraction layers, routing, and edge computing. Today, that expensive core is LLM inference.

The familiar four moves are already visible: