Ivan Nardini

AI Developer Advocate at Google Cloud. Instructor @DeepLearningAI. Startup Advisor @ycombinator. Cloud TPU, Agent Platform Engineering and AI for science.

[Ivan Nardini @ivnardini]

Presenting at the Next '26 Developer Keynote is one of those moments I'll remember for a long time.

Thank you to everyone who played a part.

Till the next Next!

[Ivan Nardini @ivnardini]

@rakyll Biased (maybe) but true

[Ivan Nardini @ivnardini]

kagent 1.0: https://t.co/9HxFvlZjT1
Agents from an MCP client: https://t.co/CPDL6UtMy8

[Ivan Nardini @ivnardini]

The Agent Substrate ecosystem is growing fast.

kagent, the open source agent framework for Kubernetes, reached 1.0 this week: its agents run on Agent Substrate, sleep when idle, and now any MCP client can discover and call them.

Resources in the 🧵

[Ivan Nardini @ivnardini]

The game to play

[Ivan Nardini @ivnardini]

Written with the Pathways on Google Cloud team. More is coming.

Blog: https://t.co/tklS4fYAQw

[Ivan Nardini @ivnardini]

Pathways on Google Cloud is now generally available

It runs JAX programs on Cloud TPUs with one Python client that lets you debug, coordinate many programs, and recover from failed hosts

I'm personally proud of this launch because I got to work with one of the great TPU engineering teams at Google

Blog in the 🧵

[Ivan Nardini @ivnardini]

GKE dynamic slicing: https://t.co/MWeRrZMxA7

[Ivan Nardini @ivnardini]

Google Cloud TPU news:

1️⃣ New TPU systems with 8i Max (long context) and 8tx (confidential inference)
2️⃣ TPU Sync, an open source KV cache offload for inference
3️⃣ Cloud Pathways in GA to run JAX with one Python process on many TPU slices
4️⃣ GKE dynamic slicing in GA to reshape TPU slices without new nodepools

Links in 🧵

[Ivan Nardini @ivnardini]

Cloud Pathways: https://t.co/7VEpLpDiPR

[Ivan Nardini @ivnardini]

If you want to know more about the new TPUs

[Google Cloud @googlecloud]

WE ARE LIVE with Google’s Chief Technologist of AI Infrastructure, @AminVahdat, to share more details on our new TPUs.

Join the broadcast now ⬇️ https://t.co/a4HiOz1mVy

[Ivan Nardini @ivnardini]

Blog: https://t.co/oBhTprb8KX

Agent Substrate on GKE docs: https://t.co/kgOaQKUdNu

Repo: https://t.co/pGu9kN3nk0

Paper (Liu et al.): https://t.co/g3zN0KsbvV

[Ivan Nardini @ivnardini]

Coding agents seem to mostly wait. In a study of agent traces by Banruo Liu and colleagues, a typical session with several prompts spent about 80% of its time waiting for the user.

Agent Substrate can snapshot the agent while it waits and give its sandbox back to a warm pool on Kubernetes. It is now 1.0, GA on GKE, and in the CNCF 🧵

[Ivan Nardini @ivnardini]

How about AX studio? 👀

[Jaana Dogan ヤナ ドガン @rakyll]

We've been thinking about building an AX Studio which makes it easier to create a local cluster and test things locally. Once tested locally, agentic tasks can be run against a production cluster without changes.

[Ivan Nardini @ivnardini]

@rakyll It was just a matter of time

[Ivan Nardini @ivnardini]

This looks interesting

[Aura @Aura_Assistant]

We're open-sourcing Autopilot.

A self-improving agent that plays your game and finds bugs. Unity, Godot and Unreal.

→ 43% more accurate
→ Almost 2X faster
→ Beats Claude Code at playtesting

It’s Computer Use for games – closing the agentic loop and solving the most tedious part of game dev.

Stop being your agent’s tester. Free trial + GitHub below.

[Ivan Nardini @ivnardini]

They also run on TPU

[Visko Platform @viskoai]

Every video you've ever watched was made before you hit play.

@willy_visko laid out what changes when it isn't. Four shifts Live Models bring 👇

[Ivan Nardini @ivnardini]

AX: https://t.co/eOmOXfz44h
Agent Substrate: https://t.co/fFhJ1ZR4EA

[Ivan Nardini @ivnardini]

I was really happy to present and discuss AX and Agent Substrate at the AGNTCon + MCPCon NA last week

With AX and Agent Substrate, coding agent runs as a session that:
- gets its own sandbox
- holds no resources while asleep
- wakes fast on a warm resources

Resources in the 🧵

[Ivan Nardini @ivnardini]

With AX, you declare each session once as a task, then start, check, or suspend it with kubectl style commands.

With Agent Substrate, the runtime keeps a suspended session as a snapshot in object storage and brings it back with its files and conversation.