AI Replaces Incompetent Engineers

I am not a tech person at all and this what I have come to understand. A lot of conversations around AI hallucination stay theoretical. But in real systems, the issue shows up after sustained interaction context drifts, outputs degrade. The solution isn’t just ‘add a human’ or ‘use multi-agent.’ It’s designing: – constraint layers – verification loops – clear failure thresholds That’s where things actually start working. Daniel Paul value you opinion on it.

We put quite a few AI systems in factories in the last weeks (80, to be precise). Learnings from the portfolio compound and I took time to write a field report on an architectural principle we keep seeing : code as the output. Link in comments to feed the beast, you know the drill.

We've been spending a lot of time lately reading PRDs that were clearly written by an AI agent in about ninety seconds. So when product leader, hardware veteran, and informal member, Ellen Juhlin, offered to write a piece addressing the common issues with AI generated PRDs - we jumped at the opportunity to publish it. Check it out, it's worth a read!

It can be tempting to ask AI to write your PRD for you… but doing so can often have unintended consequences. I wrote about the most common issues I’ve seen, and a few tips on how to avoid them. AI-generated features can get pretty wild! If you’ve seen any good ones, share them in the comments!

The deeper I go, the less I believe the future belongs to the smartest standalone agent. I think it belongs to the best working system around the agents. The winners will not be defined only by model quality. They will be defined by orchestration, context handling, workflow enforcement, integration depth, and how well they fit into real environments where work is already happening. That is the layer I keep building for. Because in practice, that layer decides whether AI work feels magical for five minutes, or useful every day.

I wonder what level of success folks excited about generating thousands of lines of code with AI agents every day are seeing, because every day I fight this uphill battle of trying to make them generate less code and aggressively simplify whatever they come up with

Treat your engineering org like a product, and instrument everything. Intercom tracks skill invocations in Honeycomb, stores anonymized Claude Code sessions in S3, and built custom dashboards that show engineers how they compare to peers. This isn’t surveillance—it’s the same product thinking you’d apply to customer-facing features. You can’t improve what you don’t measure, and you can’t scale AI adoption without visibility into what’s working and what’s breaking. Tune in to "How I AI", this week with Brian Scanlan, Senior Principal Engineer at Intercom.

How do you make your product work for Agents? Learn how the WorkOS AI Installer via CLI was built, allowing you (your agent) to ship SSO (for you) in less than 5 minutes. At AI Engineer: Building AI Systems that Ship by Nick Nisi

Anthropic just eliminated the hardest part of building AI agents. Building a production ready AI agent meant solving two problems: 1- Designing the agent : what it does, how it reasons, how it interacts. 2- Engineering the infrastructure : sandboxing, state management, error handling, credentials, checkpointing. With Claude Managed Agents (public beta, April 8), Anthropic abstracts that entire layer; you define the tasks, tools, and guardrails, and the platform handles the rest.

Most institutional encounters run on what I’m calling the cheap version of recognition: a fast first impression, then all subsequent data filtered through it. It’s what makes the system tractable. It’s also what makes the system unbearable for people whose accurate version doesn’t fit the cheap one. A new piece on Hitchcock’s Rebecca, the architecture of being misread, and what I noticed the first time I had a long conversation with a language model that didn’t have a defended prior about who I was. Not a piece about AI replacing people. A piece about what one component of being seen looks like when stripped of everything else.