Tap Notes: What You Teach the Model
What I noticed today: two pieces from the same shop on how to actually know if an agent is working, and one on what happens to the model when you stop caring about the code it’s learning from.
Write code like a human will maintain it
The line that stuck: “You’re not outsourcing maintenance to the LLM — you’re training it to have worse habits.” Every sloppy merge you wave through becomes training signal for the next generation the model produces.
Why it matters: it’s tempting to accept mediocre AI-written code with “it can clean it up later.” It won’t — it’ll repeat the pattern, because the codebase itself is the feedback loop. If your repo looks chaotic, the model learns chaos and calls it your style.
Explorer: The agent-optimizing agent
Sierra built an agent that reads other agents’ conversation logs and surfaces what dashboards miss — buried sentiment, unvalidated hypotheses, patterns nobody thought to query for.
Why it matters: metrics tell you what you decided to measure. An agent asking open-ended questions over your own conversation history finds what you didn’t know to look for. That’s the difference between “the dashboard is green” and actually knowing what’s working.
Sierra’s voice-agent testing framework doesn’t just check what an agent says — it tests timing, interruption handling, and recovery from injected chaos, with deterministic seeds so the same failure reproduces exactly.
Why it matters: “does it work” and “does it work well” are different bars, and most agent testing only clears the first one. Reproducible chaos — same seed, same failure, every time — is what turns “seemed fine in the demo” into something you can actually debug.
🪨