Everyone uses AI; few profit from it yet. Agents, systems where the model owns the control flow, are the bet the industry is making to close that gap. This week builds your mental model: the loop, the vocabulary, the autonomy dial, and the 2026 landscape.
Click each card for the “so what” and the source. Every figure is dated and linked to its source; this field moves fast, so treat the numbers as snapshots.
The length of tasks agents can complete autonomously has been doubling roughly every three months (METR, approx.). Demos get dramatically better every semester.
On long computer-use tasks, top agents score around 20% where humans score ~72% (OSWorld, approx.). Benchmarks lag the demos; reliability is the semester’s recurring villain.
One request, “find me a time with Priya next week”, handled by a chatbot, by a workflow, and by an agent. Press Play on each tab and watch where the decisions happen. The difference between the three is not intelligence; it is who decides what happens next.
Try this: run all three, then answer in one sentence each: who chose to call the calendar? who decided to stop? which run would you let act unattended, and why?