Raw

Your teams have started using coding agents this year, but the speaker's year of building AI processes reveals that classical development rules actively slow agent work and create blind spots, blocking the promised acceleration. Agent development requires restructuring teams for speed, adapting workflows for probabilistic systems, and securing the agent layer, as classical rules block acceleration and create operational risks.

**Restructure teams and roles**
- Shift to **product engineers** or **T-shaped squads** (2–3 people); product engineers build in days, classical teams stall, and T-shaped squads are the current way to move fast in uncertainty [00:10, 00:12].
- Address the **dualism of agent development**: agents need both engineering (integrations, MCP, infrastructure) and research (datasets, benchmarks, metrics); one role cannot cover both, requiring a "superhuman" or two distinct roles [00:14, 00:16].
- Plan for **3–6 months of mastery** and high resistance; invest in training and agentic development courses rather than forcing IDE bans [00:02, 00:04].

**Adapt workflows for probabilistic systems**
- Replace **Jira and human handoffs** with **agent-to-agent feedback** and **hypothesis management**; use ML System Design Docs to track experiments and communicate progress as hypotheses, not fixed features [00:06, 00:08, 00:20, 00:22].
- Keep **control points** (APIs, contracts, database) under human management while allowing agent autonomy in surrounding code; losing control of these foundations is the primary risk [00:06].

**Secure and maintain the agent layer**
- Prepare for **agent security risks**; compromised user agents can exploit MCP servers, requiring new defense strategies and entry points for this new actor [00:24].
- Assign **agent-layer maintenance** to humans; agents can unpredictably consume resources (disk, memory), so humans must monitor infrastructure and maintain the agent layer rather than just writing code [00:26].