Vicent

The web’s on fire with the OpenAI story, or at least its making the rounds on social media these last weekS. The real impact, for me, is on trust … not in AI usage as a whole but certain AI risk mitigation strategies… specially in Agentic Coding.

So what is the actual value of using one agent to check on another ? Left unreviewed I would argue that not much (ok some but not enough to trust it) … but if we have to review it … then the promise of acceleration of systems development goes out the window.

The web’s on fire with the OpenAI story, or at least its making the rounds on social media these last weekS. The real impact, for me, is on trust … not in AI usage as a whole but certain AI risk mitigation strategies… specially in Agentic Coding.

Well achtually (invoking meme) my agents do catch some stuff you say … cool but they also don’t catch other stuff and until actual expert review they are invisible as we see on many reports from Kent Beck and others.

The web’s on fire with the OpenAI story, or at least its making the rounds on social media these last weekS. The real impact, for me, is on trust … not in AI usage as a whole but certain AI risk mitigation strategies… specially in Agentic Coding.

What i have seen that they don’t really stop? Agents writing on the same tree, agents writing comments of their understanding in the code … and how is the reviewer agent (maybe using the same memory files) going to be independent from that?

The web’s on fire with the OpenAI story, or at least its making the rounds on social media these last weekS. The real impact, for me, is on trust … not in AI usage as a whole but certain AI risk mitigation strategies… specially in Agentic Coding.

I continuously audit ways to accelerate coding while ensuring quality of the system and there is a wave of solutions that are based on the concept that more then one agent can somehow avoid errors ….

The web’s on fire with the OpenAI story, or at least its making the rounds on social media these last weekS. The real impact, for me, is on trust … not in AI usage as a whole but certain AI risk mitigation strategies… specially in Agentic Coding.

Now… If a thousand agents in agreement did not catch that issue: Why does the industry believe two agents, one coding, one reviewing, or the same agent reviewing itself later, won’t make the same mistake or worse ?

The web’s on fire with the OpenAI story, or at least its making the rounds on social media these last weekS. The real impact, for me, is on trust … not in AI usage as a whole but certain AI risk mitigation strategies… specially in Agentic Coding.

Legacy code takes years to earn that name. A team disperses, memory erodes slowly until the system’s been quietly running for a decade. An agent-coded system skips all of that. It’s legacy on arrival, the day it ships, there was never a period of shared understanding on the reality of the system.

Finding out whether existing research is applicable for a new system you have in mind has never been easier. Ask an agent, it’ll find the paper (do read the paper , don’t do a “good will hunting”). Applying it correctly is a different question. Making sure it stays applied is a third one entirely.

Citing the right work is the easy part now (doing proper attribution is a must). The hard part is making sure the code written afterwards doesn’t/can’t drift from what that work says without something going red. Most systems have citations. Almost none have enforcement.

A metallurgy for agentic coding. Closed die forging: a press supplies raw force. A die shapes it. Neither one alone makes the part. Agents are the press. Cheap, enormous, undirected force. A spec is the die. It doesn’t do the work. It bounds it.

The value was never that code generation got cheap. Cheap casts with no die are just faster ways to be wrong, at scale. Skip the die and you don’t skip the legacy code problem. You speedrun it. The same undocumented assumption creeping through a system, just at agent speed instead of years.

A metallurgy for agentic coding. Closed die forging: a press supplies raw force. A die shapes it. Neither one alone makes the part. Agents are the press. Cheap, enormous, undirected force. A spec is the die. It doesn’t do the work. It bounds it.

Further still: two dies, checked against each other. Two implementations built from the same spec. Outputs are compared. Agreement is required before anything ships. Better … but agreement only proves the two pours match each other. If both misread the spec the same way, they still agree. Wrongly.

A metallurgy for agentic coding. Closed die forging: a press supplies raw force. A die shapes it. Neither one alone makes the part. Agents are the press. Cheap, enormous, undirected force. A spec is the die. It doesn’t do the work. It bounds it.

Wrong but silent output is invisible to this approach by construction. It was never looking for it. Process is fast but when something’s off, you’re just guessing. There’s nothing to hold up and inspect.

A metallurgy for agentic coding. Closed die forging: a press supplies raw force. A die shapes it. Neither one alone makes the part. Agents are the press. Cheap, enormous, undirected force. A spec is the die. It doesn’t do the work. It bounds it.

Most of the debate about agentic coding is on the spectrum regarding the die. The question that places you on it: what is the die actually checked against, if anything at all.

Really enjoyed this. The regeneration test is the cleanest cut through the spec/code debate I’ve read. But two things still feel unsolved, and I say this running a system built on exactly this architecture: the cost of deciding, and the loop.

None of this is new science, and that’s worth saying loudly. Verification, synthesis from examples, invariant mining, proof repair: university researchers built every one of these pieces decades ago, and kept refining them while the industry looked away. The pieces were never the missing part.

Really enjoyed this. The regeneration test is the cleanest cut through the spec/code debate I’ve read. But two things still feel unsolved, and I say this running a system built on exactly this architecture: the cost of deciding, and the loop.

The loop: If the system decides which production lessons become its own rules, it slowly picks the ones that make it look good. Grading its own homework. You need: today’s version judges tomorrow’s, never itself. Break rules on purpose to test the checks. The expert promotes what becomes a rule.