Raw

**Answer:** The speaker argues that the promised 10× speed‑up from coding agents is being blocked not by the agents themselves but by the need for new human roles, control mechanisms, and process changes to manage and research these agents.

**Why the acceleration stalls**

- **Agents need proper setup and learning** – Most teams defaulted to the “auto” mode of the Cursor tool, which gave poor results; the speaker found that developers require 3‑6 months to master agents and that a structured “Plan/Act” configuration is essential. [00:02‑00:04]

- **Loss of developer control creates resistance** – With code now cheap and agents generating large code volumes, engineers can no longer inspect line‑by‑line changes. The speaker cites a large marketplace that tried to put an agent in CI/CD, only to realize the agent should feed feedback to other agents, not to humans. [00:06‑00:08]

- **Both engineering and research roles are required** – Building an agent involves (a) engineering tasks (integration, MCP, control points) and (b) research tasks (datasets, benchmarks, metrics). A single “super‑human” can do both, otherwise two specialists are needed. [00:14‑00:16]

- **Process and security implications force new practices** – Traditional Agile handoffs become bottlenecks; sprints now include experiments and hypothesis testing, captured in an ML System Design Doc. Agents also introduce new security concerns (e.g., compromised shopping assistants) that require services built specifically for agents. [00:10‑00:12][00:20‑00:22]

These points give the head of engineering a clear view of which claims (training/configuration, control mechanisms, dual roles, and revised processes/security) are most relevant for the upcoming process review.