Google I/O: Jules Coding Agent, Code Mender, TPU v8 for Developers

"AGI is now on the horizon." That line from Demis Hassabis, CEO of Google DeepMind, at Google I/O 2026 was the most-quoted sentence from a keynote packed with developer announcements. It landed in context — Hassabis made the statement while emphasising the safety requirements for the agentic systems Google is now deploying at scale. The systems Google shipped at I/O 2026 are the practical argument behind his claim.

Here is everything from the developer keynote and sessions that matters to engineers building with Google's stack.

Jules: Autonomous Coding Agent Now Generally Available

Jules is Google's asynchronous coding agent, moved to general availability at I/O 2026. The distinction from every other coding tool matters: Jules does not work alongside you. It works instead of you on a discrete task.

The workflow: you assign Jules a complete task — fix a bug, implement a feature, refactor a module, write tests. Jules runs the task in an isolated virtual machine, executes the code changes autonomously, and returns a pull request ready for your review. No prompt engineering mid-task. No back-and-forth. No context switching. You assign, it delivers.

Model tiers:

• Paid plan: Gemini 3.1 Pro — the same model that scored perfect 100% on AIME 2025 with code execution
• Free tier: Gemini 3 Flash

The competitive positioning is clear. GitHub Copilot Workspace assigns agents to issues and returns diffs. Anthropic's Claude Code runs agentic coding loops in the terminal. Jules sits in the same category but is now GA with a consumer-grade product surface, not a developer preview.

For teams running large codebases with backlogs of well-defined issues: Jules is the first Google product that can clear a ticket queue without a human in the loop on each item. The PR-return model means code review stays human; execution becomes autonomous.

Code Mender: Automated Vulnerability Patching

Code Mender was not previewed before I/O. Google announced it as a surprise addition: an automated security tool that scans codebases for vulnerabilities and patches them without manual triage.

The workflow: Code Mender runs against a codebase, identifies security vulnerabilities, writes the patches, and submits them for review. No security engineer has to find the issue, understand the context, write the fix, and test it. Code Mender handles steps one through four.

The significance: the biggest bottleneck in enterprise security is not vulnerability detection — modern SAST/DAST tools are good at finding issues. The bottleneck is remediation. Security teams identify hundreds of vulnerabilities per sprint and cannot patch them fast enough. Code Mender attacks that bottleneck directly.

No pricing or availability timeline announced beyond the I/O reveal.

Managed Agents: Single API Call, Full Agent Infrastructure

Google announced Managed Agents in the Gemini API — the lowest-friction path to deploying an AI agent that has been announced by any major provider.

One API call provisions a fully functioning agent with:

• Remote sandbox (isolated Linux environment for code execution)
• Tool use and function calling
• Reasoning loop
• Infrastructure managed by Google

Previously, building a production agent required assembling orchestration, sandboxing, tool routing, error handling, and state management yourself — or using a framework like LangChain/LlamaIndex with its own complexity. Managed Agents collapses that to a single call.

Gemini Enterprise Agent Platform (renamed from Vertex AI) is the enterprise tier: unified platform for building, scaling, governing, and optimising agents at enterprise scale.

For developers currently building agent infrastructure from scratch: the Managed Agents API is worth evaluating as a replacement for custom orchestration. The tradeoff is Google infrastructure lock-in against the development time saved.

Google AI Studio: Android Apps in Minutes

Google AI Studio now supports native Android app creation. The demo showed a complete Kotlin Android application generated from a prompt in minutes — compared to weeks of setup for a new Android project from scratch.

The second capability is more significant for existing codebases: App Migration. Google AI Studio can take code from any source — legacy Java Android, web apps, other frameworks — and convert it to native Kotlin Android in hours rather than weeks. The migration handles dependency resolution, API mapping, and architecture translation.

Android Studio CLI is now stable and GA: AI agents, including Jules and external agents built on the Gemini API, can directly invoke Android Studio's build system, emulator, and test runner programmatically. Agents can write code, build, run tests, and iterate in a fully automated loop.

Firebase AI: Session Resumption and Gemma 4 on Device

Firebase AI updated at I/O with several developer-relevant changes:

Session resumption for Gemini Live API: Long-running conversations on unreliable networks can now resume mid-session without losing context. Context compression is also available. This solves a real problem for mobile apps using Gemini Live — a user dropping connectivity no longer kills the conversation state.

Grounding with Google Maps: Firebase AI can now attach real-time geospatial context to Gemini queries, reducing location-related hallucinations. Relevant for any app doing location-aware AI tasks.

Hybrid inference on iOS: On-device inference using Gemma models is now available on iOS (previously Android only). Android expanded to support Gemma 4. This enables privacy-preserving, offline-capable AI features without API calls.

Full Gemini 3.x support: Most Gemini 3.x models have graduated to GA in Firebase, ending the preview status that made production deployment risky.

TPU 8i and 8t: 80% Better Inference Per Dollar

Google announced two new tensor processing unit generations at I/O:

TPU 8i (inference-optimised): 80% better inference performance per dollar compared to the previous generation. Optimised specifically for high-concurrency agentic workloads — the exact use case that Gemini Spark and Managed Agents create.

TPU 8t (training-optimised): Released in parallel for model training workloads.

The 80% figure is the headline number. For teams running Gemini API calls at scale, the cost-per-call reduction this generates feeds directly into the pricing Google can offer on Gemini 3.5 Flash and future models. It is also the hardware explanation for how Google can run 3.2 quadrillion tokens per month at the prices it is charging.

Gemma 4: The Open-Source Story

Gemma 4 was released at I/O with Apache 2.0 licence — commercially permissive, no restrictions on enterprise use or redistribution.

Four model sizes:

• E2B and E4B (efficient small models): Native audio input for speech recognition; suitable for on-device deployment
• 26B Mixture of Experts: 6th place on Arena AI text leaderboard; efficient inference from MoE architecture
• 31B Dense: 3rd place on Arena AI text leaderboard; strongest open-weight model Google has released

All four models support native function calling, structured JSON output, system instructions, and multimodal input (video, images, variable resolutions).

The Gemma family has now been downloaded 400 million times with over 100,000 community variants. The 31B dense at 3rd on Arena is the headline — it puts an open-weight model at a capability level previously requiring proprietary API access.

For self-hosting teams, fine-tuners, and anyone with data privacy constraints that prevent API use: Gemma 4 31B is the most capable open model Google has released and one of the top three open-weight models available as of May 2026.

Project Astra: Developer Preview

Project Astra — Google DeepMind's multimodal real-world AI agent — moved to developer preview at I/O 2026. Astra can see, hear, and reason about the physical environment in real time through a camera feed.

The developer preview means external developers can now build on top of Astra's capabilities for the first time. Use cases being explored: live visual assistance, real-time object identification and interaction, physical world task automation.

No general availability date announced.

The Demis Hassabis AGI Context

The "AGI is now on the horizon" quote deserves its full context. Hassabis made the statement while describing the safety challenge of agentic AI: systems that act autonomously in the world, manage their own tool use, and execute multi-step tasks without human oversight on each step.

His argument: the capability level required to build useful agentic systems — the kind Google shipped at I/O 2026 — is the same capability level that places AGI on the near-term roadmap. The systems are not AGI. But building them requires solving the same problems AGI requires. The engineering path and the research path are now converging.

This framing matters for how Google is positioning its safety work. It is not "our models are safe enough for current use." It is "we are building toward AGI and taking safety seriously at each step." The competitive context: Anthropic said the same thing in its Constitutional AI research. OpenAI's Sam Altman has been saying it publicly since 2023.

Hassabis saying it at a public consumer keynote is the first time a Google executive has put the company in the same public framing as its frontier competitors.

Key Takeaways

• Jules GA: Google's async coding agent now generally available; assign a task, Jules runs in isolated VM, returns a pull request; Gemini 3.1 Pro on paid, Gemini 3 Flash on free; competes with GitHub Copilot Workspace and Claude Code
• Code Mender: Automated vulnerability scanning and patching; no manual triage required; attacks the remediation bottleneck in enterprise security; no pricing announced
• Managed Agents API: Single API call provisions a full agent with remote sandbox, tool use, reasoning loop, all infrastructure managed; Gemini Enterprise Agent Platform (renamed from Vertex AI) for enterprise scale
• Google AI Studio Android: Kotlin app from prompt in minutes; App Migration converts any codebase to native Kotlin in hours; Android Studio CLI stable and GA for programmatic agent access
• Firebase AI: Session resumption for Gemini Live on spotty networks; Grounding with Google Maps; Gemma 4 hybrid inference on iOS; full Gemini 3.x GA
• TPU 8i: 80% better inference performance per dollar vs prior gen; optimised for high-concurrency agentic workloads; explains Google's Gemini API pricing trajectory
• Gemma 4 31B: 3rd place on Arena AI leaderboard, Apache 2.0; 26B MoE 6th place; 400M total Gemma downloads; native function calling, multimodal, audio
• "AGI is now on the horizon": Demis Hassabis at the keynote — first time a Google executive has placed the company in the same public AGI framing as Anthropic and OpenAI
• Project Astra: Moved to developer preview; build on real-world multimodal AI agent capabilities for the first time

For the full consumer keynote recap covering Gemini 3.5 Flash, Gemini Spark, Universal Cart, Googlebooks, and Android XR glasses, read Google I/O 2026: Gemini 3.5 Flash, Antigravity, Universal Cart, Gemma 4. For how Anthropic's Claude Code compares as a coding agent, read Anthropic + PwC: 30,000 on Claude, Insurance Drops to 10 Days.