Raw

Your teams have begun using coding agents, and you are preparing for a process review. The introduction of agents has broken classical Agile handoffs, slowed delivery through waiting times, and revealed that agents require a dual engineering-research approach that Jira and traditional roles cannot support. How should we adapt our development process? We must shift from classical Agile to a hybrid model that integrates research cycles, T-shaped product engineers, and agent-centric feedback loops to manage the probabilistic nature of agent systems.

**Restructure teams to T-shaped product engineers and dual-role agents.**
- Adopt T-shaped product engineers who cover multiple roles and can build full apps in days, replacing classical Agile teams that stall for months. [00:10]
- Assign dual roles to agent development: one role for engineering (MCP, infrastructure, access rights) and one for research (datasets, benchmarks, metrics), as neither backend developers nor NLP engineers alone can deliver. [00:14]
- Prepare for junior developer friction by recognizing that agent management requires maturity and resistance to new tools can be large. [00:02]

**Integrate research cycles and hypothesis-based management into the workflow.**
- Replace Jira bug tracking for agent errors with a research cycle where errors are data collection for improving evaluation and benchmarks. [00:20]
- Use ML System Design Docs to record experiments, show clients all work, and communicate in the language of hypotheses rather than fixed features. [00:22]
- Manage sprints as a mix of features and experiments, proving hypotheses to the client rather than delivering a fixed set of features. [00:20]

**Shift from human handoffs to direct agent feedback and agent-centric design.**
- Move code review and quality checks from humans to agents, where agents review agent output directly. [00:06]
- Treat the human as an external feedback source to correct agent drift, similar to GPS correcting an aircraft. [00:08]
- Design services and security for agents as new actors, not just humans, including MCP integration and defense against compromised agents. [00:20]