Raw

**Answer:** Agents can boost development speed, but the sole barrier to the promised 10× acceleration is the human side – we must create new roles, control points, feedback loops, and process structures so that agents can operate safely and autonomously.

**1. Human resistance and skill gaps** – Teams struggle to adopt agents because developers need months to master them, resist changing IDEs, and cannot monitor the flood of generated code. The speaker notes that “developers need three to six months to become proficient” ([00:02]) and that “the price of code has dropped to zero, making line‑by‑line review impossible” ([00:04]), leading to a reliance on managers who understand the agents’ internals. A large marketplace’s failed attempt to put an agent in CI/CD, which produced redundant human‑focused comments, illustrates the flaw of treating agents like humans ([00:06]).

**2. New control points and feedback** – Trustworthy agents require clearly defined control points (contracts, APIs, databases) and continuous external feedback, otherwise they become “black boxes.” The talk cites the need to “hold certain control points” ([00:06]) and uses the aircraft‑feedback analogy to show that agents must receive corrective signals from humans or external sensors ([00:08]). Without this, agents cannot guarantee correct behavior, especially when they generate code autonomously.

**3. Process redesign – from Agile handoffs to product engineers** – Classical Agile handoffs add latency when every role works with an agent; the handoff time outweighs the speed gains of individual developers. Examples show developers finishing tasks in hours, yet “the transfer of responsibility…takes time” ([00:10]). Teams that collapse the pipeline into a single “product engineer” who owns idea‑to‑implementation achieve dramatic speed (“mobile app and website in a couple of days”) ([00:12]), though such engineers are scarce.

**4. Dual engineering + research roles and tooling** – Building agents is not just coding; it also demands research tasks (dataset collection, benchmarking, metric definition). The speaker splits an agent into (a) engineering integration (MCP, skills, ReAct loop) and (b) research (datasets, benchmarks) ([00:14]–[00:16]). Consequently, a single person cannot cover both; organizations need separate “prompt/skill designers” and “research/metrics teams,” or a rare “super‑human” who can. The ML System Design Doc is highlighted as a practical tool to capture experiments and communicate hypotheses to clients ([00:22]).

These four points together explain why the 10× acceleration has not materialized and what concrete evidence the speaker provides for each claim.