Raw

Agents can deliver the promised speed, but the real bottleneck is the human side – we must redesign roles, control points, and feedback processes before we see a 10× acceleration.  

**1. Human control points and responsibility** – The foundation of any system (contracts, APIs, databases) must stay under human oversight; otherwise agents can “run wild.” The speaker cites a failed CI/CD‑agent review experiment by a large marketplace ([00:06‑00:08]) and stresses that feedback loops must be handed to agents themselves, not to people reviewing generated code.  

**2. Dual engineering + research roles** – One specialist cannot cover both building the agent (integration, MCP, skills) and the research side (datasets, benchmarks, business‑metric methodology). The talk notes that a single “AI engineer” role is insufficient and that teams need two complementary roles or a “super‑human” who can do both ([00:14‑00:16]).  

**3. Process shift from Agile handoffs to product‑engineer model** – Classical Agile handoffs add waiting time that erodes the speed gains of agents. Product engineers who own the whole idea‑to‑implementation chain can deliver features in days, bypassing the multi‑step handoff chain ([00:10‑00:12]).  

**4. Security, integration, and new service entry points** – Agents become new actors that must interact with services, other agents, and external tools, raising fresh security and integration challenges (e.g., compromised shopping agents, need for dedicated entry points) ([00:22‑00:24]).  

These four pillars explain why the promised 10× acceleration remains unrealized and guide which claims to surface at the upcoming process review.