When an expert fixes your agent's output, where does that fix actually go?
Grand-Border3615 · reddit · 2026-10-10
An engineer building agent evals for finance/ops describes a recurring failure pattern: when accountants or compliance staff correct an agent's output, the fix lands in a Slack thread or spreadsheet, and weeks later the agent repeats the exact same mistake.
Open questions posed to the community: who decides what the right answer is, where do expert fixes go, and what happens when experts disagree? The author's takeaway: everyone is duct-taping this part of the pipeline with no systematic feedback loop.
More from coding & agent
- Decisions API hits GA as agents automate cloud video editing via feedback loops — dkundel · 2026-10-23
- Swarms API changelog: 8 months of streaming, multi-agent orchestration, and dynamic model pricing — KyeGomezB · 2026-10-23
- A curated library of 50+ free AI agent resources: videos, repos, one Google Doc — ifioknkem · 2026-10-23
- Vercel adds no-account deploys: drag and drop a file to get a live URL — steren · 2026-10-23
- Dev runs 5 Claude Code sessions that coordinate via MCP Agent Mail to root his network cameras — doodlestein · 2026-10-23
- Vercel ships 'vercel deploy --temporary' for ephemeral agent-built deployments — cramforce · 2026-10-23