I have seen the compounding teams (sundaylettersfromsam.substack.com)

Several teams are demonstrating what the author calls “compounding teams”: groups that get more than a short-lived productivity bump from off‑the‑shelf assistants by building frameworks around LLMs that let the models autonomously extend their own capabilities. Instead of vibe coding with Copilot or Claude Code alone, these teams have created proactive systems (examples include an “Amplifier”‑style framework) with callback hooks, tool calling and flow control plus higher‑level strategies so the model can decide “you’ll need a tool for that” and implement it—check it into git, add tests, and make it a permanent improvement. The result is recursive automation: tools that build tools, running many parallel processes, heavy API spend, and human attention as the new bottleneck. That pattern is already changing team structure and workflow practices. Technically, these systems lean on low‑level programmer infrastructure—filesystem access, git, Markdown, Kubernetes, XML, CI and acceptance tests—because models are surprisingly effective when they can read/write and execute through familiar dev primitives. Key implications: modular boundaries and strong testing become essential, coordination replaces routine coding as the main challenge, and the approach generalizes beyond software to many knowledge‑work tasks. There’s a steep upfront cost (months of framework work) before compounding benefits appear, but the author argues this is a clear inflection point, akin to the arrival of PCs or browsers for developer productivity.