Raw

Teams have adopted coding agents expecting acceleration, but the promised 10x gain is blocked by classical process handoffs and the probabilistic nature of agents. To unlock velocity and manage risk, the development process must shift from feature-based Agile to a model that integrates research cycles, restructures teams for full-stack speed, and treats agents as independent actors requiring new security and management patterns.

**Integrate research cycles into the delivery workflow.** Classical Agile slows down due to handoffs between roles, even when individuals work faster with agents [00:10]. Agents introduce a research dimension where errors serve as data for benchmarks rather than fixable Jira tickets [00:20]. Sprints must account for experiments and hypotheses alongside features [00:20]. Use the ML System Design Doc to record experiments and communicate hypothesis-based progress to clients [00:22].

**Restructure teams to eliminate handoffs and cover dual capabilities.** "Product engineers" who own the full process from idea to implementation deliver speed that classical teams cannot match [00:10]. Large organizations are cutting Agile teams to 2–3 T-shaped people who cover multiple roles [00:12]. Agent development requires both engineering skills (integrations, MCP) and research skills (datasets, benchmarks); since one person rarely has both, teams must include both roles or form superhuman pairs [00:14–00:16].

**Adopt agent-centric security and integration patterns.** Agents act as a new actor connecting to services and other agents, not just tools for humans; services must be designed for agent interaction [00:22]. Security risks include compromised user agents or malicious MCP servers, requiring new defense strategies [00:24]. The production process will shift toward maintaining an "agent layer" (skills, UI kits, infrastructure) while agents perform the development [00:26].

**Redefine management around control points and hypothesis metrics.** Management must focus on control points like contracts, APIs, and databases, which agents cannot yet be trusted to change, while allowing code to vary around them [00:06]. Feedback loops require human intervention as an external correction source; agents need context and periodic human checks [00:08–00:10]. Metrics must shift to business outcomes and hypothesis success rates rather than feature counts [00:22]. Use IDEF0 business functions to help analysts describe agent capabilities and simplify scope [00:18].