Fix the machine, not the output: every developer should own the AI workflow
rseroter · x · 2026-10-28
Jason Davenport argues 100% of developers should be building things that improve the workflow, not hand-patching agent outputs. Key points: manually fixing a broken build just to close the task (the "fix / really fixed" commit history) leaves the process broken; agent lineage is a SIPOC problem — if tries 1-10 were fine and try 11 failed, changing try 11's output by hand doesn't help try 12. In building cosm, agents kept reverting to stale knowledge-base behavior; adversarial prompting eventually rounded out test-case generation, and lessons had to be written back into the system. Fix the machine, not the execution output.
More from coding & agent
- Grok trading agent got stuck relaying messages between Robinhood's bot and itself — PTrubey · 2026-10-28
- Local AI Music Pipeline: LLM Subagents + YuE2 Produce a Full Eurodisco Track — Robonotes1760 · 2026-10-28
- Granting Codex or Claude sysadmin access is powerful and dangerous — RealGeneKim · 2026-10-28
- Capy moves mobile dev to the cloud with cloud-phone build and testing — k7agar · 2026-10-28
- Inside Anthropic: How Lesse and Jiang Ship Opus 5.5 and Managed Agents — Latent Space · 2026-10-28
- Replit designers shipped 765 PRs and 250K lines of code in a year — amasad · 2026-10-28