The tool-churn log: two years of developer productivity at 8am AI

If you want to know what AI development felt like between 2024 and 2026 — not the press-release version, the working-developer version — the developer-productivity thread is the receipt. It's 59 ideas of tools tried, tools dropped, things that broke, and the slow realization that the bottleneck kept moving.

The churn, in order

• Late 2024 — the field opens up. GPT-4 is the baseline; the news that Claude is now available in Canada is itself an event. The group compares coding tools, wires up N8N to connect APIs, runs OpenWebUI locally. The energy is "what can these things actually do for a build."
• Early 2025 — and the field bites back. The honest counter-current arrives. Fulvio hits AI-generated code that violates an AutoCAD license agreement; 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, methodologies for app building, the recognition that the IDE-with-an-agent is a workflow, not a gadget.
• Late 2025 — meta on the meeting itself. Even the Fireflies notetaker becomes a subject — the group is now productive enough with AI 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: collaboration and upskilling — getting a team to use the tools compatibly, not just individuals.

The three things that never stopped being true

Under the churn, the corpus keeps re-deriving the same lessons:

1. 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 Almost every tool named in 2024 was superseded by 2026.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.
2. 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 exactly why their productivity advice aged better than the hype did.
3. Individual speed ≠ team speed. The recurring frustration of 2026: one person can 10x with agents, but a team that doesn't adopt compatibly loses the gain to friction. The hard problem was never the model. It was getting humans to change how they work together.

What the churn is really measuring

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). Productivity gains kept getting eaten by the next layer up.

That's the real finding: AI made the code cheap and the coordination expensive. Two years of tool-churn was really two years of discovering that the hard part was never the part the tools were for.

Open question

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) but not yet a clean answer to the social problem on top of it.