Summary
Johan Lajili of Poolside AI — one of a small number of companies training its own foundational LLM from scratch — argues that the wildly divergent experiences people report with AI coding agents come down to a single root cause: the quality of the feedback loop available to the agent during a task.
The talk frames the divide between greenfield and brownfield codebases not as a capability problem but as a feedback problem. On greenfield projects, agents have accurate intuitions about code structure and their guesses about what will work usually do. On brownfield legacy codebases, dead ends, unused code paths, and undocumented interdependencies break those intuitions — and without a way to verify their own work, agents confidently produce wrong results and report success. The skeptical users who dismiss AI coding as overhyped are typically working in exactly this environment.
Lajili’s practical response was building a CLI tool called Spoolside that gives agents the ability to interact with a VS Code extension as if it were a web page: taking screenshots, extracting front-end and back-end logs, restarting services, and navigating the UI through high-level commands. The core principle is that an agent should reproduce a bug before claiming to fix it — and should confirm the fix before moving on. Without this loop, humans have to verify everything themselves, eliminating the autonomous operation that makes agents valuable for overnight or background tasks. Lajili closes by suggesting that engineers’ primary role is shifting toward building the scaffolding that makes AI-driven development verifiable, rather than writing product code directly.
📺 Source: AI Engineer · Published July 08, 2026
🏷️ Format: Opinion Editorial
