Using Agents to Make a Show, Not a Demo
The hard part of multi-agent work is the handoff. Get the shortest path running, and make sure it still works next time.
Contents3 sections
The Hard Part Is the Handoff
On DeepSeek Harness, the common use of agent teams is writing code. I'm using it to run a content factory: downloading, transcribing, revising scripts, and laying out pages, with the work divided among different agents. The difficulty isn't whether the model can write Chinese. The difficulty is the handoff. Where does the transcript come from? Where does the style card live? Does a draft count as published? Where does the audio go? If none of this is written down, a multi-agent setup just copies one person's verbal chaos into several people's verbal chaos.
Capabilities Are a Means
On dsh.fish, people are sharing plugins. Plugins can add many capabilities: search, image viewing, sidebars, cost tracking. My own bottleneck isn't a missing plugin. It's that after installing a lot of them, I still haven't reliably produced one episode that I can take away and deliver. Capabilities are a means. The show is the deliverable.
A Repeatable Output
So my requirements for agents are now very narrow. Don't help me build the factory into something that looks more like a factory. Help me get the shortest path working, and make sure it still works next time. Demos look good. A repeatable output is what counts as a work record.
