person signing a form with a pen

Building PiCA: Putting a Human in the Agent Loop

After working with orchestration, I started a new experiment: PiCA, an agent framework that explores approval workflows and code analysis.

Speed was not the bottleneck

Agents are already fast. What was missing, for me, was a comfortable way to stay responsible for what they do. Reviewing everything after the fact is exhausting, and approving everything without looking is dangerous. The useful design space lies between those two extremes.

Approval as architecture

It is tempting to treat human approval as a pop-up added at the end. I think it has to be part of the design. Which actions are safe to run alone? Which ones need a person? What information does that person need in order to make a real decision, rather than just clicking yes?

Lessons from operations

None of this is new to people who run production systems. Change windows, peer review and separation of duties all exist because capable professionals still make mistakes. Agents should not be held to a lower standard than the engineers around them.

Simpler every day

Experiments like this have also changed what I use daily. Today I rely on Pi, a minimal coding agent, and Herdr, a terminal multiplexer for watching several agents at work. Both are third-party tools. Building my own taught me to appreciate software that stays small and leaves me in control.

Leave a Reply

Your email address will not be published. Required fields are marked *