April 24, 2026
Banking on bots or botching banking?
Affirm Retooled for Agentic Software Development in One Week
Affirm tells 800 devs to let AI write code — internet yells “what could go wrong”
TLDR: Affirm paused engineering for a week to push AI “agents,” and now says 60% of code changes are AI-assisted. Commenters are roasting it as reckless for a financial company, predicting breaches and layoffs, while a few note human review is still in place—stakes are high when money and data move.
Affirm hit the big red button: pause normal work, unleash bots. For one week, the buy-now-pay-later giant asked 800 engineers to try “agentic AI” — think software that plans and writes code itself — and claims over 60% of code changes are now AI-assisted. They picked Claude Code as the default tool, kept work on laptops, and promised human checkpoints for judgment. The goal, they say, is to speed up shipping without breaking the money-moving machine. The crowd? Not buying the chill.
Commenters torched the move as risky for a company that literally handles cash. One set the tone with a “bets on a security breach” pool and mocked the whole thing as “vibe coding.” Another compared using AI everywhere to spraying WD-40 on everything: quick fix, sticky mess. A more technical critique warned that Affirm’s own pain points — old codebase, unstable testing, slow deploys — are exactly where AI agents struggle most. The memes came hot: “Affirm lays off 799 devs” today, “massive data breach” tomorrow. A cynic called it a last-ditch Hail Mary; another asked if they needed to read the post before calling BS. Supporters are quiet, but Affirm insists human reviews remain in place. Bold transformation or fintech faceplant? The comments are firmly in “hold my popcorn” mode.
Key Points
- •In February 2026, Affirm paused normal delivery for a week to adopt agentic AI across its engineering organization.
- •Over 800 engineers and managers were required to complete an end-to-end agentic workflow resulting in a submitted PR.
- •A working group defined a repeatable, local-first workflow using Claude Code as the default tool and human checkpoints for intent, plan approval, review, and merge.
- •Advances like Anthropic’s Opus 4.5 enabled practical coding agents that can search, plan, code, test, and iterate with minimal intervention.
- •Following the initiative, over 60% of Affirm’s pull requests became agent-assisted; the post outlines motivations, preparation, execution, issues encountered, and ongoing investments.