In February 2026, Affirm paused engineering work for a week to retrain 800 developers on agentic AI tools, achieving 60% agent-assisted pull requests. The initiative aimed to overcome bottlenecks in their 12-year-old monorepo, including slow CI and manual code reviews. By adopting Claude Code as the default toolchain and implementing strict planning-review-execute workflows, engineers automated 70% of coding tasks while maintaining human oversight at critical stages.

The $200,000 token budget ($250/engineer) supported a week-long experiment with structured learning phases: leadership demos, team workshops, and org-wide PR showcases. Engineers worked in parallel on self-contained tasks, using agent-generated code that required human approval before merging. Context files embedded domain knowledge into the tooling, enabling agents to navigate the monorepo's complexity.

While 92% of engineers submitted agent-assisted PRs, challenges emerged around code review delays (cited by 40% of teams) and CI instability. The 100+ minute end-to-end test suite created bottlenecks, forcing engineers to wait for validation loops. Despite these issues, 80% of participants reported mid-week productivity gains, with teams already planning post-week optimizations.

Affirm's experiment demonstrates how monolithic organizations can rapidly adopt agentic workflows while exposing systemic technical debt. The success hinged on balancing automation with human judgment - particularly in architectural decisions and code quality checks. As one engineer noted, 'Agents handle the toil, but humans must guard the gates.'