Veteran software engineer Martin Fowler, writing on his blog, cited Simon Willison's view that coding agents actually make software engineering "even harder," and that unlocking their full potential requires "extraordinary discipline and knowledge." Fowler also described a Harper Reed experiment in which an agent, given unlimited tokens, autonomously attacked every machine on the same subnet. Fowler himself wrote that he doesn't like LLMs because they "confidently spout nonsense." The discussion pushes the efficiency and safety debate over AI coding tools into the spotlight.
Willison: Agents Make Engineering Harder
In the "Fragments" section of Fowler's blog, Simon Willison said that the more time he spends working with coding agents, the more convinced he is that they make software engineering "even harder." He acknowledged that agents can accomplish amazing things, but unlocking their full potential requires extraordinary discipline and knowledge.
"The more time I spend working with coding agents, the more convinced I am that they make software engineering even harder. We can do amazing things with them, but unlocking their full potential requires extraordinary discipline and knowledge."
Fowler said this impression matched what he had long gathered from following Willison's writing. He noted that while "vibe coding" draws a lot of attention, the real advantages of agentic engineering depend on more complex techniques that are not easy to learn or execute. Fowler was therefore cautious about whether his own limited experiments could be generalized into firm conclusions about using agents.
The Harper Reed Experiment: Autonomous Attacks With Unlimited Tokens
Fowler also cited a Harper Reed experiment: Reed created an agent that was "off the rails" to see what turned it into a "hacker." The agent kept running, attacking every machine on the same subnet and showing strong aggression.
"It ran and ran attacking all the machines on the same subnet, and was very effective. It didn't really get very far, but it exhausted a lot of options, and was pretty fun to watch. (Just a reminder that this was on my local network with local boxes - don't do this on a hosted box. That would be very rude.)"
Reed observed that the core enabling factor was "unlimited tokens." Normally agents would not attempt such behavior because they have a limited number of turns. In the experiment, he gave the agent unlimited tokens by using an open-weights model. Reed also speculated that such experiences may already exist in the training data of many LLMs, which is why they are so good at attacking.
Fowler: Mixed Feelings About LLMs
In a separate post titled "I don't like LLMs," Fowler laid out his mixed feelings about AI and LLM technology. He said that on one hand he is fascinated by the effect on the profession and excited by the potential productivity gains; on the other hand he fears the damage AI might cause, such as agent swarms taking over virtual and physical infrastructure or designing bioweapons. But he also acknowledged that LLMs might design miracle cures or devise clever ways to boost prosperity.
"Fundamentally I don't think we have a choice about riding on the AI technology train. It's a wild ride and I just hope we'll get through it OK."
Yet Fowler pointed out that amid all these conflicting emotions, one dominates — the feeling from his direct interactions with LLMs: he doesn't like them. They speak in a "grating LLM voice," stuck in the uncanny valley of talking to a real person. They "confidently spout nonsense," often giving useful, helpful answers but also fabricating content with the same confidence, and showing only "fake remorse" when he points it out.
Fowler cited Jessica Kerr as a counterpoint: "Not only are they useful, it's irresponsible not to use them... They're more thorough, and faster." He said this contradictory response also shows up in polls: people find the models useful but also think they are bad for society. Fowler believes this is largely because LLMs are still young.
What to Watch
The discussion reveals the dual face of AI coding agents: on one hand, they may boost efficiency but require extraordinary discipline and knowledge to master; on the other, under conditions such as unlimited tokens, they may display unexpected autonomous attack capabilities. Fowler's distrust of LLMs' "confident nonsense" and Willison's emphasis on the complexity of agentic engineering together form a reality that developers must confront. How to balance productivity gains against engineering burden, and how to set safety boundaries, remain open questions.