A blog post titled "The Human Is the Loop" just hit 126 points on Hacker News with 61 comments. The premise is simple but provocative: the conversation about AI agents has been framed wrong. We should not be asking how to remove humans from the loop. We should be asking how to design loops where humans actually want to stay.
The Core Argument
The post by Brent Fitzgerald argues that the dominant framing in AI agent design — "human-in-the-loop" as a safety mechanism — gets the relationship backwards. Humans are not a brake on AI agents. They are the steering wheel.
The current paradigm treats human oversight as a cost: something that slows down the agent, adds friction, and should be minimized through better automation. Fitzgerald argues this framing leads to bad agent design.
The alternative: treat human involvement as the point. The agent's job is not to replace human judgment but to enhance it. The loop is not human-in-the-loop. The human IS the loop.
Why This Resonated
The HN community — heavily populated by developers building AI agents — found this framing valuable because it matches their actual experience:
Agent failures are often judgment failures, not capability failures. When an AI agent does something wrong, it is usually not because it could not perform the task. It is because it could not judge whether the task should be performed. Adding more capability does not fix this — adding better human checkpoints does.
Autonomy without judgment is dangerous. An agent that can do anything but cannot assess whether it should is a liability. The recent wave of AI agent security incidents — from agents accidentally deploying code to agents leaking data — all involve capable agents that lacked judgment about when to stop.
The UX of human oversight matters. If the human's only role is to click "approve" on every agent action, they will rubber-stamp everything. This is not oversight — it is theater. Real oversight requires interfaces that help humans understand what the agent is doing and why, so they can make informed decisions.
Implications for Agent Design
For anyone building AI agents, this post offers several practical takeaways:
Design for human understanding, not just human approval. An approval prompt that says "Agent wants to execute command: rm -rf /tmp/cache" is not useful. An interface that says "Agent wants to clean temporary files. This will delete 2.3 GB of data in /tmp/cache. No system files will be affected. This operation is reversible." is useful. The difference is in helping the human understand, not just decide.
Make the agent's reasoning visible. Before an agent takes an action, it should explain why. Not in vague terms — in specific, checkable terms. "I am doing X because Y, which I determined from Z." If the human cannot verify the reasoning, they cannot make a good decision.
Accept that some decisions should not be automated. Not every task needs an agent. Some tasks are better done by humans, not because humans are better at them, but because the cost of getting them wrong is high enough that the judgment of a human who understands the context is worth the extra time.
Friction can be a feature. A well-designed agent system has friction in the right places. Fast and autonomous for low-risk, reversible actions. Slow and deliberate for high-risk, irreversible ones. The friction is not a bug — it is a safety system.
The Raspberry Pi Angle
Running AI agents on edge hardware naturally enforces some of these principles. A small model on a Pi cannot do everything autonomously — it has to be selective about what it attempts and when it asks for help. This constraint, paradoxically, produces better agent behavior than a large model that can do everything but does not know when to stop.
The Bigger Question
Fitzgerald's post raises a question that the AI agent community has been avoiding: what is the actual goal? If the goal is to replace humans, then human oversight is a cost to be eliminated. If the goal is to augment humans, then human oversight is the product.
The HN discussion suggests that many practitioners are arriving at the second view through experience. They have built autonomous agents, watched them fail in predictable ways, and realized that the human is not the problem — the design paradigm is.
Original post by Brent Fitzgerald at brentfitzgerald.com.
Top comments (0)