DeepMind's Philipp Schmid Builds One Agent Three Times, Deleting Code Each Pass
AI Engineer · youtube · 2026-09-15
At AI Engineer, Google DeepMind's Philipp Schmid builds the same GitHub PR review agent three ways on stage, deleting code each pass: a hand-written Python loop, then an agent framework with auto-generated schemas — both refuse a weather question — and finally a remote sandboxed agent that just searches and answers, because nothing told it what its tools were for.
Version three runs in a hosted sandbox with a network proxy that injects credentials on outbound requests (the agent never sees the token) and an allowed-domain list; loops, tool routing, session state, and context compaction all move server-side. What remains yours: instructions, rules, capabilities as skill files, and evals.
His closing heuristic: if your harness gets more complex as models improve, you're overengineering it — some teams have replaced thousands of lines of orchestration with a few hundred lines of markdown.
More from coding & agent
- Cloudflare ships granular authz for Workers, its most requested feature — dinasaur_404 · 2026-09-15
- px0 adds millisecond full-text search, project-wide file open and auto-reload — arpit_bhayani · 2026-09-15
- Cloudflare adds granular authz to Workers with four roles for teammates and agents — dinasaur_404 · 2026-09-15
- Pick one coding agent harness, learn it deeply, and stop switching tools — iannuttall · 2026-09-15
- Hypothesis creator hiring London engineers for Hegel, open-source property-based testing — xuanalogue · 2026-09-15
- Adaption AI launches Invent API: training datasets from a prompt, no strings attached — sarahookr · 2026-09-15