Uploaded November 2026 | Updated November 2026, 2 days ago
π€ Check out Gemini Enterprise Agent Platform from Google Cloud here: https://ul.live/googlecloud_yt
π I sat down with Alan Blount (Senior Product Manager, Agent Platform, Google Cloud) to talk about what it looks like to build with agents once you stop micromanaging the steps. We totally geeked out over gardens versus factories, whether we're still engineers, one policy language from the gateway all the way down to syscalls, and why an agent's credentials should never show up in the model's context window. He also sent me off to go check out Harbor for evals.
π§ What we talk about:
π± Garden, Not Factory: Why the garden metaphor beats the software factory for building with agents, and how outcome constraints without step-by-step instructions work like fencing in the yard and then letting things grow.
ποΈ Are We Still Engineers?: What level of abstraction can we let go of and still call ourselves engineers, and why knowing what you're building and why is the line between engineering and vibe coding. Alan, a software engineer for 20 years before becoming a product manager, explains why he now defines the action space and the things he's not okay with instead of building up from the atoms.
π‘οΈ One Uber Policy Language: How the Agent Gateway pairs simple allow/deny CEL policies with a natural-language semantic policy engine, and why Alan wants a single policy language that reaches all the way down to syscalls, with sandboxes like nono working inside the container.
π Never in the Context Window: Agents acting on your credentials versus agents with their own identity, how short-lived tokens and variable substitution at the gateway keep secrets away from the model, and Alan's push for task-specific, downscoped OAuth instead of all-or-nothing access.
π Knowing What Good Looks Like: Why evals are the hard part, starting with deterministic code before reaching for a language model, and Alan's pointers to Harbor, Kaggle benchmarks and a free five-day intensive on evals.
π₯ People and links from the conversation:
π€ Alan Blount, Senior Product Manager, Agent Platform, Google Cloud, on LinkedIn: https://ul.live/ytgcalanli
π€ Alan Blount on GitHub: https://ul.live/ytgcalangh
π Advent of Agents, a daily small nugget of what you can do with agents: https://ul.live/ytgcadvent
π Kaggle, home of benchmarks and the free five-day intensive on evals Alan mentioned: https://ul.live/ytgckaggle
π§° Harbor, a framework for evaluating agents in sandboxed environments: https://ul.live/ytgcharbor
π§° nono, a sandbox for agent runtime security: https://ul.live/ytgcnono
π Agent Gateway overview (Google Cloud docs): https://ul.live/googlecloud-gateway-docs_yt
π More on this topic:
π Harness Gardens: https://ul.live/ytgcgarden
π Vibe Coding Isn't Using AI to Make Things: https://ul.live/ytgcvibe
π Building Software: Engineering Fundamentals for the Agent Era: https://ul.live/ytgcbuildsw
π From Prompt Engineering to Intent Engineering: https://ul.live/ytgcintent
π Models Are the How, Harnesses Are the What: https://ul.live/ytgcmhwhat
π The Entire Game for AI Is Articulation of Ideal State: https://ul.live/ytgcideal
π₯ The Missing Piece Is Ideal State: https://ul.live/ytidealstate
π Writing Is Closer to Thought Than Code: https://ul.live/ytgcwriting
π General Hill-Climbing: The One Main Loop in AI: https://ul.live/ytgchill
π₯ A Conversaton With Murali Rathinasamy: https://ul.live/ytgcmurali
π₯ A Conversation With Harish Peri: https://ul.live/ytgcharish
π₯ A Conversation with Duncan Greatwood: https://ul.live/ytgcduncan
π Stay connected
π¬ Subscribe to the newsletter at:
https://ul.live/ytnl
π¬ Join the UL community at:
https://ul.live/ytcommunity
π£ Follow on X:
https://ul.live/ytx
πΌ Follow on LinkedIn:
https://ul.live/ytli
π€ Check out Gemini Enterprise Agent Platform from Google Cloud here: https://ul.live/googlecloud_yt
π I sat down with Alan Blount (Senior Product Manager, Agent Platform, Google Cloud) to talk about what it looks like to build with agents once you stop micromanaging the steps. We totally geeked out over gardens versus factories, whether we're still engineers, one policy language from the gateway all the way down to syscalls, and why an agent's credentials should never show up in the model's context window. He also sent me off to go check out Harbor for evals.
π§ What we talk about:
π± Garden, Not Factory: Why the garden metaphor beats the software factory for building with agents, and how outcome constraints without step-by-step instructions work like fencing in the yard and then letting things grow.
ποΈ Are We Still Engineers?: What level of abstraction can we let go of and still call ourselves engineers, and why knowing what you're building and why is the line between engineering and vibe coding. Alan, a software engineer for 20 years before becoming a product manager, explains why he now defines the action space and the things he's not okay with instead of building up from the atoms.
π‘οΈ One Uber Policy Language: How the Agent Gateway pairs simple allow/deny CEL policies with a natural-language semantic policy engine, and why Alan wants a single policy language that reaches all the way down to syscalls, with sandboxes like nono working inside the container.
π Never in the Context Window: Agents acting on your credentials versus agents with their own identity, how short-lived tokens and variable substitution at the gateway keep secrets away from the model, and Alan's push for task-specific, downscoped OAuth instead of all-or-nothing access.
π Knowing What Good Looks Like: Why evals are the hard part, starting with deterministic code before reaching for a language model, and Alan's pointers to Harbor, Kaggle benchmarks and a free five-day intensive on evals.
π₯ People and links from the conversation:
π€ Alan Blount, Senior Product Manager, Agent Platform, Google Cloud, on LinkedIn: https://ul.live/ytgcalanli
π€ Alan Blount on GitHub: https://ul.live/ytgcalangh
π Advent of Agents, a daily small nugget of what you can do with agents: https://ul.live/ytgcadvent
π Kaggle, home of benchmarks and the free five-day intensive on evals Alan mentioned: https://ul.live/ytgckaggle
π§° Harbor, a framework for evaluating agents in sandboxed environments: https://ul.live/ytgcharbor
π§° nono, a sandbox for agent runtime security: https://ul.live/ytgcnono
π Agent Gateway overview (Google Cloud docs): https://ul.live/googlecloud-gateway-docs_yt
π More on this topic:
π Harness Gardens: https://ul.live/ytgcgarden
π Vibe Coding Isn't Using AI to Make Things: https://ul.live/ytgcvibe
π Building Software: Engineering Fundamentals for the Agent Era: https://ul.live/ytgcbuildsw
π From Prompt Engineering to Intent Engineering: https://ul.live/ytgcintent
π Models Are the How, Harnesses Are the What: https://ul.live/ytgcmhwhat
π The Entire Game for AI Is Articulation of Ideal State: https://ul.live/ytgcideal
π₯ The Missing Piece Is Ideal State: https://ul.live/ytidealstate
π Writing Is Closer to Thought Than Code: https://ul.live/ytgcwriting
π General Hill-Climbing: The One Main Loop in AI: https://ul.live/ytgchill
π₯ A Conversaton With Murali Rathinasamy: https://ul.live/ytgcmurali
π₯ A Conversation With Harish Peri: https://ul.live/ytgcharish
π₯ A Conversation with Duncan Greatwood: https://ul.live/ytgcduncan
π Stay connected
π¬ Subscribe to the newsletter at:
https://ul.live/ytnl
π¬ Join the UL community at:
https://ul.live/ytcommunity
π£ Follow on X:
https://ul.live/ytx
πΌ Follow on LinkedIn:
https://ul.live/ytli










