The tool-churn log
#8am-ai#developer-productivity#deep-dive#tooling
David Olsson
If you want to know what AI development actually felt like between 2024 and 2026 — the working-developer version, not the press release — the developer-productivity thread is the receipt. Fifty-nine ideas of tools tried, tools dropped, things that broke, and a bottleneck that kept moving.
the churn, in order
Late 2024: the field opens up. GPT-4 is the baseline. "Claude is now available in Canada" is itself an event. The group compares coding tools, wires up N8N, runs OpenWebUI locally. The energy is "what can these actually do for a build."
Early 2025: the field bites back. Fulvio hits AI-generated code that violates an AutoCAD license. The model apologizes and does it again. AI safety in coding assistance becomes a real topic, not a hypothetical. The tools are powerful and unreliable in specific, expensive ways.
Mid 2025: process over tool. The conversation matures from "which tool" to "which methodology." Pre-development planning. App-building methodology. The IDE-with-an-agent is a workflow, not a gadget.
Late 2025: meta on the meeting. Even the Fireflies notetaker becomes a subject. The group is productive enough now that the tooling around the conversation is worth optimizing.
2026: slop, healthcare, and the upskilling gap. "Slop code" gets embraced as a competitive necessity — ship fast, document the control plane. AI lands in regulated domains like healthcare documentation. And the gap that won't close: getting a team to use the tools compatibly, not just individuals.
what the churn measures
The tool you're excited about will be obsolete in a quarter. Almost every tool named in 2024 was superseded by 2026. The skill that survived wasn't tool knowledge. It was the ability to evaluate and switch.
The failures are specific, not vague. License violations. Context corruption. "You're absolutely right" as a tell. The group logged failure modes as carefully as wins, which is why this thread's advice aged better than the hype did.
Individual speed is not team speed. One person can 10x with agents. A team that doesn't adopt compatibly loses the gain to friction.
The interesting thing about a two-year tool log isn't the tools. It's that the bottleneck moved up the stack every few months. From "can the model write the code" — solved fast — to "can I trust the code" — license, context — to "can my team work this way" — still open. Each productivity gain got eaten by the next layer up.
AI made the code cheap and the coordination expensive.
If individual productivity is solved and team adoption is the bottleneck, the next tool that matters isn't a better model. It's whatever lets a group share context, conventions, and a control plane without friction. The corpus has a name for the substrate — project-state. The social problem on top of it is the next thing to build.