Raw

Your teams have started using coding agents, but the promised acceleration hasn't materialized. The bottleneck is now that classical Agile handoffs cannot keep up with agent speed, and agents introduce a probabilistic research cycle that breaks traditional sprint management. To unlock value, you must restructure teams, adopt research-oriented metrics, and treat agents as a new actor requiring dedicated oversight.

**Restructure teams toward product engineers and dual roles**
*   Shift to "product engineers" who own the full cycle from idea to implementation; classical teams stall for months without code, while product engineers can build apps in days [00:10].
*   Agents require dual roles combining engineering and research; backend developers alone build frameworks without delivering, and NLP engineers alone get bogged down in plumbing, so you need either superhumans or two distinct roles [00:14].
*   Cut Agile teams to two or three T-shaped people who cover multiple roles; large companies are already doing this to move fast in uncertainty [00:12].

**Replace classical sprint management with hypothesis-driven research cycles**
*   Agents introduce a "grey" research cycle where errors are data for improvement, not bugs; classical managers stall when faced with agent errors in Jira [00:20].
*   Sprints must include experiments and hypotheses alongside features; you must measure business metrics and talk to clients in the language of hypotheses [00:20].
*   Adopt the "ML System Design Doc" to record experiments, align with clients, and justify work; this tool helps manage the research culture required for agents [00:22].
*   Jira is convenient for humans but stalls agent work; you need to adapt tools or accept contortions to handle agent-driven workflows [00:20].

**Enforce human control points and agent security protocols**
*   Humans must remain the external feedback source; agents drift and require correction, so you cannot fully automate without human intervention [00:08].
*   Keep human control over contracts, APIs, and the database; code can change freely, but the foundation must be managed by people accountable for stability [00:06].
*   Treat agents as a new actor with security risks; services need new entry points and defenses against compromised agents interacting with MCP servers [00:22].
*   Monitor agent resource usage actively; agents can balloon disk usage or delete memory autonomously, requiring human oversight to prevent outages [00:26].