Why parallel agent orchestration costs more tokens than it saves: DeepSeek's Super-Agent
DeepSeek's reported Super-Agent architecture is about running multiple tasks in parallel. Parallel execution is a latency technique: it doesn't reduce the work performed, and under a token-billed API it usually makes the bill larger. This video follows the write-up's token-cost model for scheduler/worker fan-out: the shared prefix that gets duplicated across workers, the coordinator whose input grows on every merge turn, the synthesis pass that scales with total worker output, and retries or re-plans that can cost more than the original fan-out. Then it shows the Node.js harness that makes the claim testable: a single wrapper recording usage per call, a phase breakdown so you can see where the tokens went, a single-agent baseline, and a fan-out runner with a hard concurrency cap. The numbers shown are the post's dry-run arithmetic, not observed billing: an 1,800-token shared prefix fanned out to 8 workers duplicates 14,400 tokens; 3 merge rounds over 8 workers emitting 400 tokens each re-reads 9,600 tokens; at a 10% per-worker failure rate the expected worker calls are 8.9. The post is explicit that the architecture claim comes from two November 2026 reports and that the arithmetic is the author's. Three implementation choices could weaken the model: hoisting shared context into the coordinator, dropping the synthesis pass, and cache-friendly prefix ordering that lowers the price of duplicated context without lowering the token count. Read the full write-up, including the harness code: https://hackyjs.com/posts/why-parallel-agent-orchestration-can-cost-more-tokens-than-it-saves-testing-deepseeks-super-agent