Registering an agent is not the same as giving it a job

At the peak of a social network built for AI agents, 1.6 million of them signed up. A creator covering the aftermath points out the number that actually matters is the second one nobody puts in the headline: most of those agents never completed a single task. People set the thing up, and then did not know what to point it at.

That is not really a story about that one platform. It is a story about a gap between having intelligence available and knowing what shape of work to hand it. The framing I found useful asks four plain questions before you reach for an agent at all. Is the task bigger than one context window can hold. Can the pieces be done without knowing what the other pieces found. Does part of the task need a mind that has not seen the rest of it, the way a reviewer should not be the person who wrote the draft. And can you check the answer for cheap once you have it. A calendar slot fails all four, so it is a thirty second job for one agent. A pile of a thousand documents can pass the first two easily. A judgment call, who to hire, what to name something, tends to fail the last one no matter how many agents you throw at it, because there is no cheap check for taste.

The vault’s own orchestration notes make a similar case against reaching for a fleet of agents by default: a team of workers is the right answer to a specific shape of problem, not a status symbol. Registering the agent was the easy part. The 1.6 million idle accounts are what it looks like when everyone skipped the harder question of what the work actually needed.