Video summary
12 mistakes TELUS made building an agent platform for 100,000 employees
Main summary
Key takeaways
Summary
TELUS built Gizmos, an internal platform intended to help employees—including people with little or no development experience—safely create and share software made with AI. TELUS has more than 100,000 employees across multiple businesses and countries, which made identity, security, and deployment more complicated than a simple prototype suggested.
Gizmos provides a secure deployment environment for apps, web applications, automations, and MCP servers. It uses Cloudflare infrastructure, including Workers, containers, databases, and KV storage. The platform supports SSO, role-based access control, shared-access policies, and globally available deployments.
The 12 Mistakes and Lessons
- Making deployment seem easier than it was.
The original flow asked users to export a ZIP from an AI tool and upload it. Many users got stuck, and agents often produced projects the platform couldn’t deploy. TELUS added
llms.txtguidance and Gizmos Studio, so people can build directly on the platform.
Lesson: Observe people trying a workflow before calling it easy, and meet them where they create.
-
Leaving costs and limits until too late. Users could run expensive workloads repeatedly without understanding the cost—for example, large-scale document OCR. TELUS added accounting, usage limits, billing dashboards, and chargebacks. The speakers said Cloudflare isolates were inexpensive, while AI usage was the main cost to manage.
-
Assuming everyone used the same login provider. A Google-only login excluded employees whose business units used other identity systems. TELUS replaced the single-provider button with an email-based flow using Better Auth, which selects an appropriate login method or sends a magic link.
-
Planning for scale too late. Demos and sudden waves of users exposed bottlenecks. The team emphasized designing for growth early, especially when adoption can jump rapidly.
-
Using the main database as an analytics “closet.” Page views, deployments, and other events accumulated in the same D1 database used by the platform. TELUS moved event analytics to an analytics engine. It also found that its sampled analytics undercounted people; moving user counts back to D1 produced a figure about seven times higher.
Lesson: Decide what questions you need to answer before choosing where to store the data.
-
Letting a community marketplace become cluttered. AI makes it easy to create many apps, including duplicates and unused projects. Gizmos added organizational and team views, usage metrics, and stars to help people find relevant tools. The aim is to distinguish what users can see from what they need for their work.
-
Operating without adequate observability or a status page. Initially, users often reported failures, and the team couldn’t easily tell whether the problem was with Gizmos or the user. Adding tracing itself disrupted outbound calls. TELUS later created a status page on a separate Worker with a five-minute scheduled check, so it could remain available during a platform outage.
Lesson: Publish a status page from day one.
-
Chasing frameworks and rewriting too often. The speakers cautioned against repeatedly changing frameworks, runtimes, or tools. AI can make rewrites feel cheap, but most tools can achieve the desired result; avoid rebuilding simply because a new option is interesting.
-
Assuming everyone works in a terminal or uses the same AI tool. The team initially expected users to deploy through Claude Code, but many worked in Claude’s desktop or browser experiences and wanted to avoid the terminal. TELUS adapted by making Gizmos usable from those environments and adding a Claude connector via MCP.
Lesson: Ask where users actually work before prescribing a workflow.
-
Risking the creation of yet another platform. Gizmos aims to support many kinds of tools and workloads rather than solve only one narrow problem. The presenters acknowledged the danger of creating another platform that duplicates existing ones; their goal is to make Gizmos a platform for deploying and managing other software.
-
Allowing AI-generated code to accumulate without refactoring. Rapid “vibe coding” led to duplicated implementations, layers of wrappers, and excessive comments. A dedicated refactoring effort consolidated repeated approaches; the team reported removing about 18,000 lines of comments in one day and nearly 100,000 lines overall. They still use AI to write code, but recommend scheduling refactoring before the codebase becomes unwieldy.
-
Trusting AI output without understanding the system. Agents can produce code and documentation quickly, but generated diagrams or project documents can create a false sense of understanding. Developers still need to decide how the system should behave, test it, and understand the tools and architecture themselves.
Security, Governance, and Platform Details
In the Q&A, the presenters described isolating each Gizmo on its own domain, using authentication redirects and access controls to keep apps separate, and scanning submissions for secrets and potential problems with a Gemini-based check. Gizmos Studio can help users address flagged issues and redeploy. They advised against putting sensitive data on the platform until the relevant protections and data-privacy requirements are in place; TELUS also has data-residency obligations.
The platform supports both isolates for many I/O-oriented workloads and containers for more CPU-intensive work, such as data analysis. “Custom bindings” let one Gizmo expose functionality to another under a permissions mechanism. The team is also working on agent identities and delegated permissions.
The presenters reported that Gizmos had reached roughly 15,000 users and nearly half a million views in the prior month. They described it as a safe place to deploy and manage the software AI agents are increasingly able to create.
Main Speakers and Source
- Kevin Ng — TELUS developer advocacy; discussed user workflows, adoption, and platform lessons.
- Andrew Culberson — TELUS software engineer; discussed infrastructure, scaling, and integrations.
- Justin Watts — Mentioned as the engineer who led a major refactoring effort; the opening speaker appears to introduce the talk.
- Source: A TELUS engineering presentation published by Cloudflare Developers.
Rate this summary
Your feedback will help improve summaries.
Improve this summary
Reprocess with a stronger model when the summary feels incomplete or inaccurate.
Translate summary in another language
Ask questions to this video
Chat for follow-up questions, clarifications, and source-backed answers.