@aakashgupta: DeepSeek gave an AI agent a team of helpers and 20 jobs to get done, and in five tries it handed off exactly zero of th…

DeepSeek gave an AI agent a team of helpers and 20 jobs to get done, and in five tries it handed off exactly zero of them. Turns out AI has the same problem as every founder who can't let go of the code. The setup was a group chat with 300 messages, about 20 real work requests buried inside, and a tool that let the agent pass tasks to helper agents. It tried to do every job itself, and every new message in the chat interrupted it mid-task. The slowest jobs took almost half an hour. So the team rewrote the prompt and basically begged it to delegate. It listened in four of five runs. In the fifth it decided the work was "small" and did it anyway, and its slowest tasks went right back up to 29 minutes. DeepSeek's fix is the same one good companies eventually land on, which is taking the keyboard away from the boss. The top layer, called the Router, reads every message from chat and GitHub and decides who does what, and it runs without a terminal, so it literally can't do the work itself. Each task goes to its own Worker with full tools and a clean slate, and the Worker reports back to the Router. With the boss off the keyboard, the slowest tasks finished in under 16 minutes. The pay structure flips too. The Router does fine on a lighter, cheaper model because sorting messages is easy, and the expensive model goes to the Workers writing the code. So the manager ends up being the cheapest employee in the building. Once the boss stops doing the work, it starts noticing things. In one of DeepSeek's examples, someone in a public user group complained that clicking a session in their Mac app kept triggering a rename, dozens of times a day. Nobody tagged the agent. It reproduced the bug in the background, traced it to Electron swallowing the first click on an unfocused window so the next click counted as a double-click, and opened a fix with before and after screenshots. The load test is where it gets weird. When DeepSeek cranked incoming requests from 1 a minute up to 40 a minute, the Router's time per decision dropped from 6.4 seconds to 1.6, because requests that stack up during one step get handled together in the next. Give a human manager 40 times the inbox and they fall apart, and this one got 4x faster. None of it is free. The two-layer setup burned 23.2M cached tokens per run against 6.1M for the single agent, and the test used simulated coding work, so treat the numbers as directional. DeepSeek estimates a person can keep track of 3 to 5 agent sessions before losing the plot. The whole design spends tokens to buy back human attention, and right now tokens are by far the cheaper of the two.