Channel
Interviewed Person
Various
Modern copilots can now build for you. More people than ever can submit code to production. How can you maintain the performance of your platform by treating observability data as machine-first input? We still need dashboard for humans – but we also need logs for agents that automatically patch errors, tune performance, and enforce security. We’ll explore how Vercel can weave logs, metrics, traces, feature-flag telemetry, and experiment results into a closed feedback loop where AI agents both watch and repair production systems – so you can keep everything performant and stable – when anyone can build.
[Music] Hello everyone. Uh my name is Matt and I am a technical consultant at Versel. Um and my role at Versel has been pretty fascinating at for the last year. It's been a cross-section of um seeing many companies evolve uh to the recent progress in AI and tooling. And I thought I'd want to talk a little bit about one of those customers and the transformation they're going through and
uh also provide some key takeaways and uh of keeping up your code quality in this new way of working. Um Versel if you don't know it it's a developer it's a a cloud hosting platform but it uh has two sides of the product. We have the developer experience platform and the managed infrastructure. And my job is to regularly work with around 20 of our enterprise customers. Um, and I really just focus on team productivity and uh app performance. So in this last year,
we've had a crazy amount of AI tools coming through and it's been really interesting to see each company adapt um to the AI transformation and evolve with the new team uh new tools that are available. Um the company that I want to talk to you today um talk to you about today is uh Teada Pharmaceutical. Uh this company is a Japanese pharmaceutical firm that is 220 years old. Um they are a global company who have mandated AI
transformation throughout the company from the CTO to the shop floor. Uh they've built their own AI model gateways. They have an internal agent marketplace amongst their uh amongst their teams uh where teams can kind of create their own agents with the secured data of the firm. And recently they've started to play with triggering agents from observability alerts. So this is me trying to visualize their plan. Uh if we can follow it here. Um
this is deploys per day. Um the product team in blue, this is how many they're hoping to deploy. And then they want the whole company in yellow to start deploying um using various tools which are now available. And uh then the big thing that we're going to talk about today is the agent side of things. So the green deploys coming up through. This is this is um me just trying to show you what it would look like if their plan was successful. um they're looking for an explosion in in contributors um for humans that humans that are technical, non-humans and uh um
sorry humans that are technical, non-technical humans and AI agents. Um, and this means I've been quite closely working with their product team. And uh, this is kind of the the plan for the next six months is to keep that product team steadily rising, have the nontechnical contributors go up and then have a huge increase from the number of agents that are directly deploying on Kada. Um, and then this is the next year after
that. But um I I took some liberties on that slide. But what I'm trying to say is that um none of their PRs are kind of like automatically deploying today, but they're already trying to be kind of bleeding edge of uh what's possible. And um without human intervention, uh they're getting pretty close at just being able to do it automatically. And what they're building increasingly feels like months and not years away. So last week actually this usual quote
came up but uh Mike Quer anthropic has just said that um 70% of claw codes PRs are created by claw code itself. Uh incident have a plan next year to have completely um kind of a aentic uh incident solving agents that have just done for you. Um and I'm going to go a little bit of step by step how do we get there? How did Ticada manage to kind of get to the position where they're trying to be the bleeding edge? Uh the current way. So imagine you have a uh a core web vital