Today I gave a presentation on the future of GenAI. Nowadays, I always include one slightly controversial take: that humans might someday play second fiddle to autonomous agents.
It is too easy to assume that only humans will delegate to agents, and not the other way around.
Human-agent interaction will take forms we can barely imagine today.
Peter Gostev recently joked about ProxyHumans on LinkedIn, and I found it amusing, as I can easily imagine it becoming reality.
As engineers, we can already feel this shift. Instead of picking up a single ticket and grinding it to completion with one coding agent at your side, you can pick up three tickets, spawn three coding agents (plus a few background or sub-agents), and round-robin feedback until all of them are ready for a pull request. This pattern was described here by the Anthropic team. Coaching several agents that run in parallel while you context-switch, nudge, and course-correct transforms you into something new: A herder of agents.
I suspect this pattern will extend far beyond software engineering. Marketeers might launch ten copywriting agents. Product managers could spin up fleets of research bots. Today, many people have already experienced this by opening three ChatGPT tabs and asking each a different question, then doing a round of follow-up questions.
Getting value from this setup is simple in theory but maddening in practice: you lose track of threads, forget which answer came from where, and context-switch until your biological RAM thrashes. My current approach is a digital scratchpad where I can stack dump after every interaction. When I return to a thread, it helps to reload the mental stack.
The future may still debate whether humans are subordinate to agents, but before that argument is settled, we will all become expert herders.