Speakers: Drew Breunig (Cmpnd)
Track: Keynote
AI agents learn what we want from our feedback, then compress that intent into prompts, skills and memories—and discard the evidence. The result is a collection of natural-language artefacts that decay as models change and edge cases accumulate. Prompts rot; skills bloat; memories quietly encode stale assumptions.
Drew Breunig argues that feedback and execution traces should be treated as source code. Instead of preserving only prose instructions, teams should retain the examples that express their intent and repeatedly compile them into task-specific workflows. He will show why smarter models do not remove the need for this discipline, how code and other precise representations can replace fragile instructions, and why compiled workflows can be more auditable, reliable and maintainable than general agents—especially inside enterprises.
https://data.webdirections.org/ai-engineer-sydney/sessions.json#e68972a1-4a4d-4114-94c1-945d59cd2b68
Full text on webdirections.org
.