You Didn't Ship a Bug. You Just Wrote It for a Human.
summary
A demo agent that calls Gmail Calendar and a handful of connectors runs on a 15‑minute schedule and repeatedly strains the production DB; latency spikes trace to one last seen timestamp that was updating 60 times a second until batching fixes it. The uploader claims 60% of LLM errors stem from rate limits. The concrete takeaways: built an agent that worked but broke human‑centered infra, fixed the symptom with batching, and argues for fine‑grained agent identities, just‑in‑time scopes, and visibility into each action.
AI-written summary · about
chapters (7)
- 0:00Human-focused architecture limits
- 1:03Agents expose auth flaws
- 3:10Agents break permission models
- 6:38Agents vs human code
- 8:44Fine-grained agent permissions
- 10:50Guardrails for agent auth
- 12:13Rethink agent architecture
related talks
30mWhy Agents Don’t Ship: Closing the Production Gap for Real Business Impact
37mYour Software is Broken — Pay Attention
51mActually shipping agents and not just talking about it
21mRemote Agents, or: How I Fixed Our Page's Performance Without Really Doing Anything
25mAgents are not Apps: Stop building interfaces, Start building coworkers
32mWhat Agents Do When You’re Not Looking
19mYour Biggest User Isn't Human: Designing DevTools for AI Agents
19mAgents in Action: How Google is Automating Operational Toil