@googledevs: Most agent demos are stateless. Ask → answer → context gone. Production agents need to run for days — pausing, resuming…

Most agent demos are stateless. Ask → answer → context gone.

Production agents need to run for days — pausing, resuming, persisting state.

@addyosmani joins @smithakolan on the Agent Factory to build three real agents with ADK 2.0 and Gemini 3.5 Flash, including one that runs Doom in the browser.

Check out this episode and more in the Agent Factory series → https://goo.gle/3QdbgMw

TL;DR

Software engineers can use AI agents like Anti-gravity, leveraging skills, MCP, and clear architecture review to boost productivity while maintaining control over code quality. Agents not only accelerate coding but also handle everything from prototyping to daily drudgery.

How Agents Are Changing the Way Engineers Work

In a conversation with Rohde Davis, head of Google Developer Tools team, we explored how AI agents help developers achieve “10x or even 100x productivity.” Davis—a former top-100 open source contributor on GitHub who can program in 18 languages—shared how agents fit into his daily workflow.

Reducing Cognitive Load, Accelerating Prototyping and Learning

For Davis, the core value of AI lies in reducing cognitive load. Examples:

• He now writes richer, more comprehensive test suites than he could manually, enabling confident releases.
• He uses agents to build first, second, and third prototypes—and even quickly switch to a different framework or language.
• For personal drudgery (e.g., building a marketing site), he has agents generate pages, take screenshots, and integrate content.

“AI is accelerating everything—the learning process, the building process, research and ideation.”

Daily Work Setup: Skills and MCP Are Core

Davis’s daily workflow revolves around skills and MCP (Model Context Protocol). He explains that skills are essentially cheat sheets for agents—compressed context that integrates multiple scripts, reusable type resources, design systems, and serves as triggerable workflows.

Skill Examples

• Obsidian skill: for note-taking and building knowledge bases.
• Flutter or Firebase skill: for apps needing data sync.
• Global skills: Android CLI, Chrome extensions, DevTools, modern web guides (CSS methods, etc.).

The Value of MCP Servers

In multiple projects, Davis enables MCP for Flutter and Dart. It allows hot reload and use of developer tools, providing build support for Anti-gravity. He also adds project-specific skills (e.g., a skill for the shadcn design system) so the model can explore how to use them.

“When you give the coding agent the right context, the quality and output of generated code changes dramatically.”

Code Review Strategy: Architecture First, API Contracts Are Lifeline

Asked how much agent-generated code he reviews, Davis said he has reviewed more code in the past six months than the previous few years—because he’s producing much more. But his review has clear priorities:

• Architecture level: He looks at how parts integrate, likening it to pruning a bonsai—constantly trimming to keep it concise but not simple. He’s moving toward a flat architecture: state separated from UI, UI separated from data. This lets him quickly spot when an agent goes off track (e.g., files in unexpected places).
• API contracts: For backends, he “cares very much about the interface” because even a small schema change increases cognitive load on users. Once the API and contracts pass review and tests, he confidently lets agents auto-generate the rest.
• Example first: He likes to write a single example for the agent, which then replicates the pattern and fills in the rest, reducing guesswork.

“If you’re building a backend, any small change to the API schema greatly increases the cognitive burden on users to update and fetch different versions.”

Evolution of Anti-gravity 2.0: From Bundled Experience to Independent Components

Davis walked through key updates to Anti-gravity at I/O: separating the IDE from the agent manager. Previously a bundled experience, it now has independent CLI, IDE, and SDK that can be installed separately. For example, users unfamiliar with the IDE can use the CLI directly, even via SSH into a server.

He also showed a multi-agent parallel demo: via a voice prompt (“build a multilingual note-taking app with Vite, Markdown, URL sharing, Go + SQLite backend, and generate a marketing website”), Anti-gravity produces an implementation plan and spawns sub-agents. Voice input already includes programming language and product knowledge, outperforming some generic speech-to-text systems.

Practical Applications: Offline Recommendations & Personal Tools

Davis’s personal website heavily uses Gemma (Gemma 4 local model). He uses Anti-gravity to generate Gemma summaries for each article, vectorizes them with Gemma embeddings, and runs a generator at compile time—enabling article recommendations for the current view without any server, fully offline and dynamically based on content.

He also built a tag-based time-box manager (grouped by “today”, “last week”, etc.). Anti-gravity built the extension in Swift from just a prompt, and he quickly ported it to a Chrome extension, reducing the search overhead when switching between Git projects.

The Future of Engineering: Understand the Underlying & Embrace Agents

Davis emphasized he still loves writing code—especially studying old hardware or writing Go. He believes understanding how interfaces and standard libraries work is essential, because knowledge from the “research, prototype, build” phases transfers to future projects. Agents handle repetitive work, freeing engineers to focus more on design and review.

“Skills are a cool way to compress context for the model. I like using skills to integrate multiple scripts, reusable type resources, and design systems.”

Source: Video: “Most agent demos are stateless…” from @googledevs (https://www.youtube.com/watch?v=Dk4MD6TNiWE&list=PLIivdWyY5sqLXR1eSkiM5bE6pFlXC-OSs)