Raw

**Answer:** The speaker argues that the expected 10× speed‑up from coding agents has not materialised because the real bottleneck is human involvement and the lack of new processes, roles, feedback loops, and security practices needed to make agents trustworthy and effective.

**Why this matters**

- **Human resistance & loss of control** – Teams default to the “auto” mode of agents without understanding them, leading to a three‑to‑six‑month learning curve and reluctance to give agents authority over critical contracts, APIs and databases. The speaker cites a real‑life example of a large marketplace trying (and failing) to let an agent review code in CI/CD, showing that agents cannot yet be trusted with control points. [00:02‑00:06]

- **Process & role redesign** – Classical Agile handoffs become slower when every role works with an agent; the handoff time outweighs the speed gains in coding. Successful “product engineers” who own idea‑to‑implementation achieve rapid delivery, but such talent is scarce. Agent development also demands two distinct roles: an engineering side (integration, MCP, skills) and a research side (datasets, benchmarks, metrics). One person cannot reliably cover both. [00:10‑00:16]

- **Feedback loops are essential** – Agents must receive continuous external feedback (e.g., human‑provided context, sensor‑style data) to stay aligned; otherwise they cannot self‑correct. The speaker likens this to aircraft autopilot systems that rely on external GPS corrections. [00:08‑00:10]

- **Security & future organisational shift** – Agents become new actors that need dedicated entry points, access‑right controls, and protection against compromise (e.g., a shopping‑assistant agent hijacked). Existing tools like Jira are ill‑suited for managing agent‑generated work; the human role will evolve into maintaining the “agent layer,” overseeing context, skills, and infrastructure, while research and experimentation become routine parts of development. [00:22‑00:26]