control / note 10

I can't trust it to run alone.

The real question with agents: how far should one run before you look?

Delegation is not amplification.

Delegation asks the system to own the outcome. Amplification keeps the outcome yours while the machine supplies scale and stamina. This sounds like philosophy until something goes wrong and you find out which one you built.

Errors compound in the dark.

A chain that runs unwatched multiplies its small mistakes. The safe pattern is boring: show the inputs, show the outputs, make the next step obvious, and let a person decide whether to continue.

What to build.

Put checkpoints where the consequences are: drafts before sends, previews before publishes, queues before deletes. Make the stop control obvious and instant. Keep a log a person can actually read. Let it run free on anything reversible, and prefer many small runs with a person choosing between them over one long chain nobody watched. For the rest, the machine proposes and the user decides.

design principle

Keep goals, checkpoints, and accountability with the user; let the machine bring the scale.

one implementation

Pachinko's version: Feeds, Functions, and MCP clients all move work forward, but everything they do lands as visible notes and queued runs you can inspect, redirect, or delete. Our feeds store history. And while Functions can be chained, the default designs for manual, instead of automated long chains.