AI agents: what can they actually do?
From a good answer to a finished task. A practical look at the capabilities, the limits, and the human decisions in between.
18 stories
Analysis & guides
Explanations, practical guides, and a closer look at the research.
Copilot canvases make the workflow a shared interface
GitHub’s walkthrough shows an agent building a board around the task.
Local AI: where to start
Choose the task, the hardware, and the operating boundaries before choosing the model.
A multilingual model can still miss the culture
Cohere researchers trace how cultural diversity narrows through training data.
VoiceEQ looks at what speech benchmarks miss
Hume AI researchers examine the human quality of voice interaction.
ScarfBench tests agents on Java migrations
IBM evaluates modernization across a project rather than an isolated function.
Control over agents is a system, not just a prompt
Google DeepMind describes layered oversight for increasingly capable internal agents.
Voice AI depends on the plumbing, too
OpenAI explains changes to the infrastructure behind real-time conversation.
DeepSeek V4 asks what long context is actually for
Hugging Face examines Pro, Flash, and the needs of long-running agents.
Command-line tools now have two kinds of users
Mistral’s Spaces engineering story considers interfaces for people and agents.
EVA evaluates the conversation, not just the transcript
ServiceNow researchers look at both task completion and the human experience of voice AI.
Measuring intelligence takes more than one score
Google DeepMind proposes a cognitive framework for evaluating AI progress.
A Rails testing agent needs to run what it writes
Mistral engineers describe a specialized loop for generating and checking tests.
Claude’s constitution makes its intended values visible
Anthropic publishes a revised account of how it wants the model to behave.
A year after DeepSeek, the open ecosystem is wider
Hugging Face looks beyond one breakout model to a changing community.
AI at work: measure tasks before predicting jobs
Anthropic’s Economic Index adds a more detailed view of how people use Claude.
Beyond the parameter count
A new architecture paper is a reminder that bigger is only one way a model can change.
How to read an AI benchmark
Start the year with a better question: what, exactly, did the test measure?