Raw

Your teams have started using coding agents this year, but the promised acceleration is stalled because classical Agile handoffs and role definitions clash with how agents work, and agents introduce a research dimension that breaks standard sprint planning. To unlock value and avoid the pitfalls the speaker encountered, we must shift from classical Agile to a model that treats agents as a new actor requiring research cycles, T-shaped product engineers, and explicit control points, while redefining human roles toward boundary exploration and agent-layer maintenance.

**Restructure teams toward T-shaped product engineers and dual-role agent builders.**
*   Classical teams slow down; "product engineers" who take the whole process can build apps in days, not months [00:10].
*   Large companies are cutting Agile teams to two or three T-shaped people [00:10].
*   Agents require two roles simultaneously: engineering (integrations, MCP, infrastructure) and research (datasets, benchmarks, metrics); one person is rare, so you need two or a superhuman [00:14–00:16].
*   Backend devs build frameworks instead of agents; NLP engineers do plumbing; analysts struggle to describe agents without IDEF0 mapping [00:14–00:18].

**Introduce research cycles and hypothesis-based management into the SDLC.**
*   Developing agents creates a "research process" inside the classical scheme; sprints must include experiments and hypotheses, not just features [00:20].
*   Use the ML System Design Doc to record experiments and align with the client on what worked and what didn't [00:22].
*   Adopt a "language of hypotheses" with the business: "We will test this many hypotheses; something will work, something will not" [00:22].
*   Jira is convenient for humans but stalls agent development when errors are logged as bugs; engineers don't know how to code Jira bugs one by one [00:20].

**Establish explicit control points and treat agents as a new actor.**
*   Code cost has dropped, but developers must hold control points: contracts, APIs, and the database; agents can "dance" around these, but losing control here is fatal [00:06].
*   Agents should review agents; feedback loops must fall on the agent's role, with the human outside checking results like unit tests [00:08].
*   Agents are a new actor connecting to services; services need new entry points and security models for agent-to-agent interaction [00:24].
*   A large marketplace tried an agent-in-CD review, but it piles comments on humans unnecessarily; the human is the external source for feedback, like GPS correcting a plane [00:08].

**Shift human work to boundary exploration and agent-layer maintenance.**
*   The only thing blocking 10x acceleration is the human; agents cope well, but humans must find their new place [00:26].
*   Humans become researchers and maintain the "agent layer": developing UI kits, skills, and watching infrastructure (e.g., disk usage, memory) [00:26].
*   Product owners throw ideas into the agent layer; QA still answers for quality, but the main task is maintaining the layer that does the developing [00:26].
*   A product owner's agent ran out of disk and deleted its own memory; this requires human oversight [00:26].