I have been thinking about how AI agents work, especially this idea of replica agents versus first principles ones. It comes up a lot in discussions like on Hacker News. The question is basically, when should these agents just copy what humans do in their workflows, and when should they kind of ignore all that and go their own way.
A lot of the agent designs out there start by asking how a human would handle the task. Like, they would open up a browser, click on some buttons, follow these standard operating procedures or SOPs. That approach seems pretty safe, I guess, because it is familiar. But it also feels limiting, like it is not pushing things forward enough.
One way to frame it that I have seen is replica agents are good in situations where humans need to check on things, step in if needed, or just trust what is happening step by step. On the other hand, first principles agents might be better when the main goal is just getting the most efficient outcome, without all the human habits getting in the way.
The tricky part is figuring out where to draw that line. I am not totally sure myself. Like, at what point do you stop trying to automate the worker exactly as they are, and instead start solving the actual problem in a direct way. It seems like some people might have heuristics for deciding this, maybe based on the risks involved or the type of task. That is the part that stands out to me, how to make that call without overcomplicating it.
Sometimes it feels like copying human stuff is easier at first, but then you get stuck. Other times, going from first principles could miss some practical details that humans handle intuitively. I think there needs to be more on that balance. ~
I’ve been writing about building Agent-First SaaS and working with teams implementing LangGraph flows.
I’ve noticed a recurring pattern where we get stuck trying to perfectly replicate a human's SOP (e.g., "click this button, then read this PDF"). While reproducing human workflows is great for trust and "human-on-the-loop" auditing, I argue it often traps us in a local optimum.
This post explores the difference between "Replica Agents" (biomimicry) and "First-Principles Agents" (optimizing for the objective function). I draw on examples like Amazon's "Chaos Storage" and AlphaGo to suggest that sometimes the most efficient agent workflow looks nothing like the human one.
Curious to hear how others are balancing "legibility" vs. "efficiency" in their agent designs.
I have been thinking about how AI agents work, especially this idea of replica agents versus first principles ones. It comes up a lot in discussions like on Hacker News. The question is basically, when should these agents just copy what humans do in their workflows, and when should they kind of ignore all that and go their own way.
A lot of the agent designs out there start by asking how a human would handle the task. Like, they would open up a browser, click on some buttons, follow these standard operating procedures or SOPs. That approach seems pretty safe, I guess, because it is familiar. But it also feels limiting, like it is not pushing things forward enough.
One way to frame it that I have seen is replica agents are good in situations where humans need to check on things, step in if needed, or just trust what is happening step by step. On the other hand, first principles agents might be better when the main goal is just getting the most efficient outcome, without all the human habits getting in the way.
The tricky part is figuring out where to draw that line. I am not totally sure myself. Like, at what point do you stop trying to automate the worker exactly as they are, and instead start solving the actual problem in a direct way. It seems like some people might have heuristics for deciding this, maybe based on the risks involved or the type of task. That is the part that stands out to me, how to make that call without overcomplicating it.
Sometimes it feels like copying human stuff is easier at first, but then you get stuck. Other times, going from first principles could miss some practical details that humans handle intuitively. I think there needs to be more on that balance. ~
I’ve been writing about building Agent-First SaaS and working with teams implementing LangGraph flows.
I’ve noticed a recurring pattern where we get stuck trying to perfectly replicate a human's SOP (e.g., "click this button, then read this PDF"). While reproducing human workflows is great for trust and "human-on-the-loop" auditing, I argue it often traps us in a local optimum.
This post explores the difference between "Replica Agents" (biomimicry) and "First-Principles Agents" (optimizing for the objective function). I draw on examples like Amazon's "Chaos Storage" and AlphaGo to suggest that sometimes the most efficient agent workflow looks nothing like the human one.
Curious to hear how others are balancing "legibility" vs. "efficiency" in their agent designs.