Point it at a website, tell it which fields you want, get back clean structured records. That's the agent we're designing in this post — and the interesting part isn't the model, it's the harness decisions that make it actually reliable at scale.
Same model, same weights, zero retraining — LangChain changed nothing but the scaffolding around a coding model and jumped it from 30th place to the top five on a benchmark. That scaffolding has a name: the harness. And it's the part you actually control.
"The model is the engine — but the harness is everything else." In Episode 7, we dig into why the infrastructure layer around your AI model matters more than the model itself, rank the best models available right now, and ask whether the open-weighted revolution is about to make frontier subscriptions obsolete.
More instruction, worse output. Zyte's head of R&D on why telling your agent exactly what to do can blind it to the obvious answer.
Is GLM-5.2 really closing the gap to Anthropic - and at just a fraction of the cost - or is it just more AI hype? I think so, and let me show you why.
"Four people, four diets, two work schedules, and a baby who answers to nobody. That's what finally made me build a personal agent." A walkthrough of the actual architecture I run to hold my household and my DevRel work together — profiles, skills, memory, and the web-data layer that makes it all reach the live web.
AI agents can generate code, suggest selectors, and draft crawl logic. What they can't do is design the system that decides when to stop, what to trust, and how to recover when the web pushes back. That job still belongs to a human.
Multi-agent orchestration is having its moment. The diagrams are everywhere now. Boxes for planners, boxes for hands, boxes for daemons, arrows to a shared brain, a human floating at the top. They keep getting prettier. The part where the web pushes back is still the part nobody draws.
Data-gathering doesn’t have to be memory-intensive. You can fit the world’s weather on a 9cm-square board, when you move the work to a web scraping API.
For the last 30 days, I did one thing almost exclusively: I built scraping systems with AI agents, from the ground up, across real targets, with real deadlines. Not prototypes designed to impress in a demo, not isolated experiments running against a toy website, but production-grade pipelines that needed to ship and keep running.
I've been running a series of conversations with developers at Zyte to understand what's actually changed in the way they work since LLMs showed up. Not the headlines. The day-to-day. What they delegate, what they don't, what they notice, what surprises them.
This one was different on two counts.
New models can process larger inputs, and confuse themselves in the process. Context management techniques can solve the problem.
_HFpro5d6k3.png&w=256&q=75)
_E4PyVpfAxa.png&w=256&q=75)
-(1).png&w=1920&q=75)
-(1)_VZGHqxCgXV.png&w=1920&q=75)
