Raw

Your teams started using coding agents this year. The speaker, who took over an AI engineering unit a year ago, reports that the promised 10x acceleration has not materialized; instead, teams face resistance, handoff bottlenecks, and a mismatch between classical Agile and the probabilistic nature of agents, requiring new roles and feedback mechanisms. What specific changes does the speaker recommend, and what evidence supports them? Agent development demands a shift from classical Agile to a hypothesis-driven model led by product engineers who manage feedback and infrastructure, because agents introduce a research cycle and new security risks that traditional processes cannot handle.

**Process and workflow.** Move from feature-based sprints to hypothesis-driven work and use the ML System Design Doc to record experiments and align with clients. The speaker found that classical Jira workflows stall agent development because errors are data for benchmarks, not bugs, and managers struggle to report on "ten experiments" per sprint; the ML System Design Doc solved this by making research visible to the client [00:20], [00:22]. For analysts, replace agent descriptions with IDEF0 business function mapping to clarify control points and integrations; this moved stalled teams off dead points and helped cut an agent with 100 tools down to essential functions [00:16], [00:18].

**Roles and team structure.** Replace classical Agile teams with product engineers or T-shaped squads of two to three people who cover multiple roles and build features rapidly. The speaker observed that backend developers waste months building frameworks instead of agents, and NLP engineers get bogged down in plumbing while product managers misunderstand the backlog; in contrast, product engineers assemble apps in days where classical teams take months, and large companies are cutting teams down to T-shaped units to survive uncertainty [00:10], [00:12], [00:14]. Humans must shift to maintaining the agent layer—managing UI kits, skills, and infrastructure—because agents can autonomously consume resources or delete memory, as seen when an open-source assistant ran out of disk and wiped its own memory [00:26].

**Feedback and control.** Establish external human feedback loops and route automated reviews back to agents rather than humans. The speaker argues that agents cannot exist without constant feedback on code execution, browser behavior, and errors, and humans act as the external correction source, illustrated by an underfloor heating example where the agent lacks context about the user's state [00:08], [00:12]. Automated CI/CD reviews should feed results to the coding agents, not humans; the speaker cites a large marketplace's flawed attempt to have an agent review human developers, noting that since agents write the code, the review loop must close on the agent [00:06].