"I burned all my tokens researching how to save tokens."

Quesma's post-mortem on a custom deep-research pipeline whose token-optimization effort cost more than it saved. 155 points, 197 comments. The value is in the specifics of where multi-agent research pipelines leak: redundant fan-out, re-fetching the same sources across agents, and context re-injection between stages. I run a pipeline shaped exactly like this and all three leaks are real. Re-fetching across agents is the worst one because it looks like parallelism.