Affirm Hit 60% Agent-Assisted PRs in Four Months. The Playbook Is Public.

The most concrete data point on enterprise AI coding adoption this week came from Affirm's engineering blog. A 9-person working group retooled their entire org for agentic development. Not in a quarter. In one week. Four months later: 60% of pull requests are agent-assisted, up from near zero. Weekly merged PR volume up 58% year-over-year.

The specifics matter here. Affirm standardized on Claude Code as the single default tool. Local-first development, not cloud sandboxes. Explicit human checkpoints at exactly four points: intent, plan approval, code review, merge. Everything between those checkpoints, the agent handles. That's a specific, replicable division of labor, and it's the first time I've seen a public fintech company publish the playbook this clearly.

The 58% throughput increase is the number skeptics should sit with. Affirm isn't a startup demo. They process real financial transactions with compliance requirements, code review standards, and uptime obligations. If their engineers are shipping 58% more merged PRs with agent assistance, that's production evidence, not a benchmark.

My take: the one-week retooling timeline is what makes this replicable. Most engineering orgs overthink the rollout. Form a committee. Run a pilot. Evaluate for a quarter. Write a strategy doc. Affirm's approach was blunt. Pick one tool. Make it the default. Define where humans must be in the loop. Ship. The 60% number didn't come from mandates. It came from making the agentic path the path of least resistance.

This connects directly to the enterprise SaaS crash in Story 5. When a fintech company can boost throughput 58% by giving engineers agents, the market starts asking: what else can agents replace?