When agents go from demo to production, it's the tooling around the model that breaks
Embarrassed-Milk4818 · reddit · 2026-10-27
A Reddit discussion thread: the author's agent worked perfectly in demos but broke quickly in production—and the failures were in the stack around the model, not the model itself. Typical breaks include tool calls that look fine but use wrong data with no clear error, silent token expiry, retrieval steps quietly returning wrong chunks, an 85%-reliable step becoming unreliable when chained eight deep, and untraced handoffs between agents or tools. The author asks others for one concrete break and one concrete fix: approval gates for risky actions, tool idempotency, full execution tracing, or narrowing the toolset.
More from coding & agent
- 5 Must-Bookmark Sites for Grok Bot Builders: Marketplace, Guides, Docs, Templates & METAGROK — victor_explore · 2026-10-27
- Speculation: Rockstar gave the whole GTA 6 team unlimited Claude access — theteknosaur · 2026-10-27
- Typosquatting nearly cost him his Keychain — he reverse-engineered the malware with an AI agent (50-min writeup) — rchardkovacs · 2026-10-27
- A leaked prompt turns Claude Code into a self-building skills machine — nikola_mr64990 · 2026-10-27
- Open-source Agent Skills review research papers claim by claim against their evidence — paulnovosad · 2026-10-27
- Reddit debate: do AI agents remove work or just create a review queue? — Few-Garlic2725 · 2026-10-27