GenAI PMInteractions API
A new API for executing code and managing agent memory in Google’s hosted sandbox workflow. It matters to AI PMs as part of the control plane for agent execution.
Key Highlights
- Interactions API evolved from a beta unified interface for Gemini models and agents into a tool for hosted code execution and memory management.
- Google I/O demos positioned it as part of the control plane for secure agent execution in a hosted Linux sandbox.
- Improved error messages made the API more readable for both developers and autonomous agents.
- Persistent container IDs suggest support for longer-lived agent workflows, with cold-start latency as a current tradeoff.
Interactions API
Overview
Interactions API is Google DeepMind’s emerging interface for running agentic workflows that combine Gemini model calls with hosted tool execution, including secure code execution and agent memory management. Based on newsletter mentions, it has evolved from a beta unified interface for Gemini models and agents into a control-plane layer for Gemini Managed Agents, where developers can provision a hosted Linux sandbox and manage execution state through a single API-driven workflow.For AI Product Managers, this matters because it points to a more opinionated platform approach to agent development: instead of stitching together model inference, sandboxing, memory, and orchestration manually, teams can rely on a managed API surface. That can reduce engineering complexity, speed up prototyping, and create clearer product tradeoffs around persistence, latency, observability, and safety when shipping agent features.
Key Developments
- 2026-01-03: Google DeepMind introduced the Interactions API in beta, described by Phil Schmid as a unified interface for Gemini models and agents, positioned to support agentic features in 2026.
- 2026-05-05: Logan Kilpatrick announced major error message improvements for the Interactions API, making responses more readable for both humans and agents.
- 2026-05-05: Philipp Schmid highlighted a quality-of-life upgrade: errors now specify the exact invalid field and value, list supported enum options, compare expected vs. actual formats, and identify field paths such as `input[0].name`.
- 2026-05-23: At Google I/O, Philipp Schmid demoed building an AI agent with its own secure, hosted Linux sandbox in a single API call using Gemini Managed Agents and the Interactions API to execute code and manage memory.
- 2026-05-23: Additional clarification noted that each sandbox session receives its own persistent container ID, which can be shared across agents, with the current tradeoff being stronger persistence at the cost of slower cold starts.
Relevance to AI PMs
1. Faster agent feature delivery: Interactions API suggests a managed way to bundle model access, execution, and memory into one workflow. PMs can use this to shorten time-to-MVP for coding agents, research agents, or workflow assistants without designing every infrastructure layer from scratch. 2. Clearer product tradeoff decisions: The mentions around persistent container IDs and slower cold starts are highly relevant for roadmap planning. PMs need to decide when persistence improves user value enough to justify latency, especially for multi-step tasks, debugging flows, or long-lived agent sessions. 3. Better developer and agent UX: Improved error messages are not just a developer convenience; they reduce integration friction, speed up debugging, and make autonomous retries more reliable. PMs evaluating platform adoption should treat error quality and API ergonomics as core product criteria, not minor implementation details.Related
- google-deepmind: The organization behind the Interactions API and the broader Gemini ecosystem in which it is being positioned.
- gemini: Interactions API was initially framed as a unified interface for Gemini models and agents, making Gemini the foundational model layer.
- gemini-managed-agents: The most direct adjacent product mentioned alongside Interactions API; together they enable hosted agent execution with secure sandboxing and memory management.
- phil-schmid / philipp-schmid: Repeatedly cited as a key public explainer of the product, including the beta announcement and Google I/O demo details.
- logan-kilpatrick: Associated with shipping usability improvements, especially better error messages, which signal active iteration on developer experience.
Newsletter Mentions (3)
“Philipp Schmid at Google I/O demoed how to build an AI agent with its own secure, hosted Linux sandbox in a single API call using Gemini Managed Agents and the new Interactions API to execute code and manage its memory.”
#8 𝕏 Philipp Schmid at Google I/O demoed how to build an AI agent with its own secure, hosted Linux sandbox in a single API call using Gemini Managed Agents and the new Interactions API to execute code and manage its memory. #9 𝕏 Philipp Schmid clarifies that each sandbox session gets its own persistent container ID—shareable across agents—and that they’re currently trading persistence for slower cold starts.
“#8 𝕏 Logan Kilpatrick rolled out extensive error message improvements for the Interactions API, making its feedback far more human- and agent-readable.”
#7 📝 OpenAI News How OpenAI delivers low-latency voice AI at scale - OpenAI explains engineering techniques to achieve low-latency voice AI at scale, covering system design, model optimizations, and infrastructure approaches. The post outlines how these measures reduce end-to-end latency for real-time voice applications. #8 𝕏 Logan Kilpatrick rolled out extensive error message improvements for the Interactions API, making its feedback far more human- and agent-readable. #9 𝕏 Philipp Schmid launched a QoL upgrade for the Interactions API: errors now name the exact field and bad value, list supported enum options, show expected vs. actual formats, and pinpoint field paths like `input[0].name`. #10 𝕏 Guillermo Rauch launched npx deepspec, an open-source agent orchestrator that leverages thousands of parallel coding agents in Vercel Sandbox to uncover critical security vulnerabilities in minutes.
“GoogleDeepMind Interactions API in beta : Phil Schmid @_philschmid announced the Interactions API , a unified interface for Gemini models and agents , set to power agentic features in 2026.”
Google DeepMind Announces Interactions API Beta From X AI Product Launches & Updates GoogleDeepMind Interactions API in beta : Phil Schmid @_philschmid announced the Interactions API , a unified interface for Gemini models and agents , set to power agentic features in 2026. Qwen-Image-2512 7s high-res image generation : Qwen @Alibaba_Qwen showcased high-resolution images in ~7s from Qwen-Image-2512 , crediting @PrunaAI for support. Agent Workflows integration with ACP : Llama Index @llama_index unveiled an integration of Agent Workflows with the Agent Client Protocol (ACP) , enabling fully customizable agentic systems.
Related
AI engineer and builder who shares implementation examples for real-time translation apps. He is referenced twice for Gemini Live-based app work.
Google’s AI research and model organization. The newsletter cites its AI Control Roadmap as a framework for managing advanced AI behavior.
A product leader known for AI developer platform commentary. In this newsletter he is quoted on Google AI Studio and a free platform that speeds app creation.
Gemini is Google’s AI model family used here for generating workout video imagery and content prompts. The newsletter references a specific Nano Banana model within Gemini.
AI product and developer advocate who shares predictions on generative AI trends. Relevant for AI PMs tracking market direction and product strategy.
Stay updated on Interactions API
Get curated AI PM insights delivered daily — covering this and 1,000+ other sources.
Subscribe Free