Redeploy reposted this
Just read this article from Armin Ronacher that really resonated with me https://lnkd.in/dmschA2i Working in a large-ish system with many contributors has honestly become a massive pain with agents and the pace of slop you have to wade through. That shared language that existed before is largely ignored by agents, and I see a ton of solutions that sidestep already-established and proven patterns in a repository, obviously 100% vibed with little thought or attempts at understanding what already exists. The code is vibed, the PR description is not written by the person who opened the PR and in common agent fashion contains everything you don't need to know about the change in plain text. The friction that used to force you to actually learn a system before changing it is gone, and with it goes the understanding that friction used to build. The agent will definitely try to get that understanding, but it will very likely often fail. Insufficient grounding, a too-small view of the system, or knowledge that simply doesn't exist in any file it can read. So it does the confident thing instead of the correct one. At Babel, losing that shared language stopped construction, but with agents, construction doesn't stop. The code compiles, the tests pass, the changes keep landing, and the explanations get generated on demand; all while the architectural understanding that would let humans reason about the system together disappears. It feels like building a Jenga tower, and it feels like this gets proven again and again with production failures caused by agents. Yes, most likely misuse, but see again my point about organizations not being ready before they hand out access. Adding domain-specific skills to the repo feels like a bandaid. Sure, you can spend time and build an organization-wide MCP server like Stripe's Toolshed, but will you do that before it is too late? Do you have the pre-existing infrastructure and knowledge to do it? Agents have massively outrun the pace at which many organizations are ready to actually operate. Agentic tooling is being handed out left, right and center without necessarily thinking about the consequences nor preparing for those consequences. Increased speed does not equal increased efficiency or quality, in a majority of cases I've encountered. Add more people to a project and you fall victim to Brooks's Law, add agents and you fall victim to more changes landing that no one synced on nor necessarily fully understand. In my experience guardrails have become more important than ever, but it is almost impossible to make them robust enough, and I'm starting to think they might be the wrong layer to solve this at anyway. If the real thing we're losing is shared human understanding, no amount of tooling at the code layer restores that. I think mindset is one solution, i.e. you must take full ownership of what lands on main. Unfortunately, that is rare. Anyone else feel this pain?