An agent is a system that runs a loop. It perceives state, reasons with a model, acts through tools, observes the result, and goes again, in pursuit of a goal that outlives a single turn. Everything below sits somewhere on the three axes. The interesting design decisions are always about where.
Where agents demonstrably work at scale. Terminal, IDE and async.
The race for the front door. Someone else runs the loop.
OpenClaw and its children. The runtime shrinks; the ownership doesn't.
The web is being quietly redesigned to be clicked by machines.
Where the procurement budget actually goes.
The build-your-own tier. Loops, graphs, crews and handoffs.
The three organs everyone is reinventing, plus somewhere safe to run.
The agentic web's OAuth moment, still being argued about.
Depth over breadth. Priced on outcomes, not seats.
Both directions. The SOC tier and the offensive tier.
The interface a persistent agent probably wants.
Hypothesis, experiment, analysis. Some of it in Nature.
The quieter corners of arXiv, plus one uncovered petri dish.
The leg of the race the text-window discourse misses.
Not agents. Engines. Reviewing these as though they were cars is the category error.