Raw

We must restructure our process to support a research cycle, dual roles, and agent-centric design, because agents introduce probabilistic behavior that breaks classical Agile handoffs and Jira workflows.

Your teams have started using coding agents, but the expected acceleration has not materialized. Classical Agile handoffs now slow progress, and agents introduce a probabilistic research loop that Jira cannot manage. To move forward, we must restructure our process to support a research cycle, dual roles, and agent-centric design, because agents introduce probabilistic behavior that breaks classical Agile handoffs and Jira workflows.

**Restructure teams into product engineers and split dual roles**
- Classical teams stall for months; product engineers build apps in days, creating a "mad gap" in velocity. [00:10]
- Large companies are cutting Agile teams to two or three T-shaped people who cover more roles. [00:10]
- Agents require dual roles (engineering and research); one person is a "superhuman," otherwise you need two. [00:16]
- Backend developers build frameworks instead of agents, and NLP engineers do plumbing; neither fits alone. [00:14]

**Institute a research cycle and hypothesis tracking**
- Jira stalls agent development because errors become data for improvement, not bugs to fix one by one. [00:18]
- Sprints must include experiments and hypotheses, not just features, to prove value. [00:20]
- The ML System Design Doc records experiments and aligns clients on hypothesis outcomes. [00:22]
- Agents need feedback loops where the human acts as an external source to correct drift. [00:08]

**Define agents via business functions and prepare services for the agent actor**
- IDEF0 methodology helps analysts describe agents as business functions, moving teams off dead points. [00:16]
- Breaking agents into functions reveals excess tools and simplifies the agent's ideology. [00:18]
- Agents are a new actor connecting to services; services must be ready for agent entry points and security. [00:22]
- Compromised agents pose risks like unauthorized shopping; services need defense mechanisms. [00:24]
- Agents can autonomously consume resources; humans must maintain the agent layer to prevent disk/memory exhaustion. [00:26]