DeepSeek's Super-agent: A Router Above Your Workers, Not Superintelligence
The bottleneck in AI-assisted work is human attention, not model capability. DeepSeek's architecture trades extra tokens to offload the "who's doing what" bookkeeping to the system, letting one person oversee a team's backlog instead of juggling windows.
DeepSeek's Super-agent prototype isn't a new superintelligence; the name just means the supervisor sits above the workers. The architecture has two layers: a persistent Router keeps watch over chat, GitHub, and email, decides whether an incoming message needs action, and hands it to a dedicated Worker. Workers are isolated, each handling one task, while the Router holds the global picture and speaks for the system, so six workers don't spam the group chat at once.
A "subscription direct-pass" rule sends follow-up comments on a GitHub PR straight to the Worker that already owns that PR, skipping the Router's re-dispatch. Removing it raised Router decision calls by about 30% and tokens accordingly. In an internal test, the system lurked in a team group and, unprompted, reproduced and fixed a double-click bug after a human complained about it, posting a PR with before-and-after screenshots. Hard prompt rules cap replies at 100 words, block "I'll handle it" filler, and limit PR responses to three rounds to stop AI politeness loops.
Benchmarks put median task completion at 12.4 minutes and the 95th percentile at 15.7, versus 29.5 at the 95th percentile for a single-session-plus-subagent-tools baseline. Remarkably, in five runs the baseline model never once delegated to its available subagents, even with prompts telling it to. DeepSeek's answer is to make division of labor structural, not a model choice. The cost is roughly four times the read-cache tokens, a trade-off the team frames as spending tokens for shorter, steadier completion times.
Making division of labor architectural instead of leaving it to model judgment is the core bet; the five-run baseline that never delegated shows why that decision can't be trusted to an LLM.
Four times the tokens buys predictable completion times, but it's only worth the expense when the human supervisor is the constraint; a one-off low-priority bug doesn't justify it.
The hard reply limits reflect a real failure mode: an agent with initiative generates noise, and noise trains people to ignore it, so the system is deliberately made annoying-proof.