Raw

Teams have adopted coding agents, but the expected acceleration is blocked by classical process friction and the unique nature of agent development. To capture value, we must shift to hypothesis-driven cycles, reorganize around product engineers with dual capabilities, and redesign services for agent interaction.

**Adopt hypothesis-driven cycles and research documentation**
*   Classical Agile and Jira stall agent development because agent errors represent data for improvement, not bugs to fix; treating them as bugs causes teams to stall [00:20].
*   We must introduce and measure business metrics, communicating with clients in the language of hypotheses rather than feature counts [00:22].
*   Use the ML System Design Doc to record all experiments, making the research work visible to clients and justifying decisions [00:22].

**Reorganize around product engineers and dual-role capabilities**
*   Handoffs between analysts, developers, and product people create waiting time that negates agent speed; product engineers who own the full process deliver incredible speed [00:10].
*   Agents require a dual role combining engineering (integrations, MCP, infrastructure) and research (datasets, benchmarks, metrics); one person cannot usually hold both, so we need either superhumans or pairs [00:16].
*   Shift from large Agile teams to T-shaped teams of two or three people who cover multiple roles to move fast in uncertainty [00:12].

**Redesign services and control points for agent actors**
*   Agents are a new actor connecting to services and other agents; our services are not ready for this, requiring new entry points, interaction models, and security defenses against compromised agents [00:24].
*   Maintain human control at critical points—contracts, APIs, and databases—while allowing agents autonomy over internal code and UI generation [00:06].
*   Implement feedback loops where the human acts as an external correction source; agents cannot exist without constant feedback on execution and errors [00:08].