AI Challenges in Software Development

REWORK host Kimberly Rhodes asks 37signals co-founders Jason Fried and David Heinemeier Hansson how AI changed Basecamp 5's development. They describe faster design exploration and production diagnosis, but also more pressure on architecture, review, product scope, and token economics.

Key Points Covered

• Most work now starts with a prompt: Hansson says most Basecamp fixes and upgrades begin with agents, while designers use working betas to discover what they actually want [00:02:06]-[00:05:21].
• Compliance is not judgment: Agents can implement unnecessary complexity without challenging the request, leaving performance, security, and architectural problems for experienced reviewers [00:06:08]-[00:10:21].
• Prototype code is disposable: The team may preserve the validated design, HTML, and CSS while replacing unsuitable Ruby or JavaScript before merge [00:08:16]-[00:11:19].
• Diagnosis is a strong use case: Hansson reports useful agent loops over logs, unusual data shapes, queries, and measured optimizations, while broad feature work still needs tighter steering [00:12:23]-[00:13:26].
• More capacity needs more restraint: Fried and Hansson warn that a larger menu of feasible features can bloat a product rather than improve it [00:14:31]-[00:18:30].
• Output and value diverge: More generated code or features do not guarantee proportional revenue, while agent costs can continue to rise [00:27:31]-[00:32:51].

These are 37signals' first-party observations, not a controlled comparison of productivity, quality, or cost.

Full video: https://www.youtube.com/watch?v=pdPXbXE5xw0(opens in a new tab)