AI Revisited

REWORK host Kimberly Rhodes speaks with 37signals co-founder and CTO David Heinemeier Hansson about his move from AI skepticism to regular terminal-agent use. Hansson values agents that execute, test, and investigate independently, while preserving expert review and the freedom to discard their output.

Key Points Covered

• Agent mode fits differently from autocomplete: Terminal agents can follow a plan, use shell tools, and work independently without interrupting the developer's own typing [00:00:51]-[00:04:57].
• Useful drafts still need ownership: Hansson often keeps substantial generated work after iteration or finishes and rewrites it himself [00:04:57]-[00:06:54].
• Agents can reduce operational toil: 37signals applies them to security-report triage and investigations across logs and monitoring systems; Hansson also identifies routine access-log review as a candidate use [00:06:54]-[00:13:18].
• Debugging can combine unfamiliar tools: One agent tests hypotheses around a Rails console defect, locates the responsible change, and proposes a temporary patch [00:15:26]-[00:18:32].
• Parallel drafts are inputs, not answers: Five models produced working MCP connector drafts, but Hansson retained none unchanged because they missed his interface goals [00:18:32]-[00:21:19].
• Criticality sets the boundary: Non-programmers gain new creative power, but systems involving sensitive data or consequential failure require stronger supervision [00:25:34]-[00:30:43].

The success rates and retained-code estimates are personal reports; the episode does not measure failed sessions, review cost, or security accuracy.

Full video: https://www.youtube.com/watch?v=AkeTmfUW4_U(opens in a new tab)