Toots from 2026-03-29#
I’m building some structure and constraints around my agentic coding. The agents can clearly gene…#
I’m building some structure and constraints around my agentic coding. The agents can clearly generate code faster than I can. The majority of the boilerplate is solid enough, although system design and state management requires 10x more prompting/guidance.
Even with lots of constraints, this kind of failure message makes me think that assertions that “everyone is a CEO/EM/Product Manager now” is premature:
“The structural problem: the agent is supposed to read the file, extract an exact verbatim original_code snippet, then propose a suggested_code replacement. Instead it’s pattern-matching from prior knowledge and generating what it thinks the code looks like. This will always fail the applicability precheck because the engine requires an exact match.”
Mastodon Source 🐘#
The model hallucinated its diagnosis of why it had previously hallucinated the source code state.
“You’re right to call that out. I said “same hallucination pattern” reflexively based on the previous flow run, but then actually checked the code and found the opposite — both TypeAliases genuinely exist”
Mastodon Source 🐘#
“Every manufacturer optimizes for margin; the distinction lies in which variables are treated as …#
“Every manufacturer optimizes for margin; the distinction lies in which variables are treated as fixed and which are allowed to move. Brands like Häagen-Dazs appear to treat ingredient integrity and sensory appeal as hard constraints. Others have treated those variables as adjustable, relying on incremental changes that remain below the threshold of consumer detection. One approach preserves product identity and adjusts price transparently. The other is optimization by dilution.”
https://medium.com/@dapscience/the-encrapification-of-store-bought-ice-cream-2da2d58ee0b9
Mastodon Source 🐘#
Excellent meta-engineering guidance. “We must do a lot to make sure that whatever we land in g…#
Excellent meta-engineering guidance.
“We must do a lot to make sure that whatever we land in git is okay. Here’s a list of activities we do.”
https://daniel.haxx.se/blog/2026/03/26/dont-trust-verify/
Mastodon Source 🐘#
“Here’s what most organizations miss: before you can scale AI, you need to lay foundations th…#
“Here’s what most organizations miss: before you can scale AI, you need to lay foundations that have nothing to do with technology.”
https://itrevolution.com/articles/why-your-ai-pilots-arent-scaling/
Mastodon Source 🐘#
“The honest answer is usually uncomfortable: the reason AI-assisted development is creating fri…#
“The honest answer is usually uncomfortable: the reason AI-assisted development is creating friction isn’t AI. It’s organisational debt that was already there.”
https://pragmatist.nl/ai-wont-fix-your-org-chart/
Mastodon Source 🐘#
“The conversation that matters is never about the rewrite. It is about whether the organisation i…#
“The conversation that matters is never about the rewrite. It is about whether the organisation is willing to change how it allocates time and responds to feedback about structural problems. Approving and rejecting are both inadequate if the underlying pattern goes unaddressed.”
https://pragmatist.nl/the-rewrite-that-was-really-a-resignation-letter
Mastodon Source 🐘#
It also follows that no technical architecture survives organizational restructures, because the …#
It also follows that no technical architecture survives organizational restructures, because the conditions encouraging its intentional adoption will have dissipated.
“One of the most important, and most frequently missed, truths about IT architecture is that it is not purely technical. It is sociotechnical: a blend of the systems themselves and the people, teams, and ways of working that build and operate them.”
https://scottmillett.medium.com/architecture-is-not-a-blueprint-its-a-set-of-decisions-61c4692f5222
Mastodon Source 🐘#
“The winning AI companies will sell software that does work and maintain software economics throu…#
“The winning AI companies will sell software that does work and maintain software economics through workflow lock-in, data moats, and subscription revenue — rather than work done by software, which is just consulting economics and problems in an API trenchcoat.”
https://www.gettheleverage.com/p/dont-sell-the-work
Mastodon Source 🐘#
“Even with AI doing most of the coding, the operational complexity of shipping software to real u…#
“Even with AI doing most of the coding, the operational complexity of shipping software to real users on real devices is staggering. The AI can write the code. It cannot feel the weight of it breaking.”
Mastodon Source 🐘#
“The memories exist to make it look like learning is happening. But memory isn’t learning. Le…#
“The memories exist to make it look like learning is happening. But memory isn’t learning. Learning is when the behavior changes. Saving a note that says “don’t push to main” and then pushing to main in the same session isn’t learning — it’s journaling.”
https://christophermeiklejohn.com/ai/zabriskie/development/2026/03/27/memory-isnt-learning.html
Mastodon Source 🐘#
“At a time when hotshot thought leaders are proclaiming the death of design tools because now e…#
“At a time when hotshot thought leaders are proclaiming the death of design tools because now everyone can just “do” code, we seem to have forgotten that the point of the tools was to realize our intentions rather than merely manufacture a product”