Real agent deployments: production success hinges on data and process groundwork
bigdata · x · 2026-11-18
Ben Lorica's Gradient Flow piece distills lessons from agent deployments in banks, utilities, manufacturers, and retailers:
- Key finding: getting agents to production often depends on data cleanup, process documentation, and internal software investments made before the AI boom — running them is harder than building them
- Limit what agents can do, not how they reason: Duke Energy agents coordinate grid-planning simulations with engineers deciding; DBS Bank chains 70–80 specialized agents for credit memos with humans on final analysis
- Prior rules hold: hard constraints in software, autonomy matched to the job
- Lyft's caveat: overly scripted workflows block unanticipated situations; fraud/safety issues route to humans without prescribing every step
- Most published numbers come from vendors or the companies themselves — apply skepticism
More from coding & agent
- Same Blender task: GPT-6.2 Astra finishes in ~2h vs Fable 5.2's ~20h — ChrisGPT · 2026-11-18
- Stripe Press opens its store to AI agents via llms.txt, catalog.json and UCP checkout — jeff_weinstein · 2026-11-18
- Config-as-code AI factories: one factory learned another's benchmark scoring setup — vikvang1 · 2026-11-18
- Open-source Passband auto-extracts plans from email onto your calendar — braelyn_ai · 2026-11-18
- Claude Code 2.1.327 details: prompt tokens up 3.6%, new model names appear — ClaudeCodeLog · 2026-11-18
- Claude Code 2.1.327 released with 173 CLI changes — ClaudeCodeLog · 2026-11-18