managed agents
A managed-agent infrastructure pattern with self-hosted sandboxes mentioned in Google DeepMind's launches. It reflects the growing emphasis on secure, controllable agent execution environments.
Key Highlights
- Managed agents separate model reasoning from action execution, improving control, security, and scalability.
- Anthropic framed the pattern as decoupling the “brain” from the “hands” in large-scale agent systems.
- Google’s Managed Agents added remote sandboxes, background tasks, MCP, function calling, and later self-hosted sandboxes.
- The concept is especially relevant for AI PMs shipping agentic workflows with code execution, network access, or enterprise integrations.
At a glance
- Category
- Agent Platform
Managed agents
Overview
Managed agents are an infrastructure pattern for running AI agents in controlled execution environments, where the model's decision-making is separated from the systems that actually perform actions. Across the mentions here, the concept shows up in both Anthropic and Google ecosystems: Anthropic framed it as “decoupling the brain from the hands,” while Google DeepMind and the Gemini API emphasized dedicated remote sandboxes, self-hosted sandboxes, background execution, MCP connectivity, function calling, persistent filesystems, and networked tool access.For AI Product Managers, managed agents matter because they turn agentic workflows from a prototype into an operable product surface. Instead of letting an agent run directly in a fragile or over-permissioned environment, managed-agent architectures add security boundaries, controllability, and scalability. That makes them especially relevant for products involving long-running tasks, code execution, enterprise integrations, regulated workflows, or any roadmap where agent reliability and governance are as important as model quality.
Key Developments
- 2026-04-21: Anthropic Engineering published Scaling Managed Agents: Decoupling the brain from the hands, outlining architectural principles for scaling agent systems by separating reasoning from execution.
- 2026-05-06: Claude launched financial-services agent templates for tasks like pitch building, valuation reviews, and month-end book closes, available in Cowork and Claude Code, or deployable via cookbooks as Managed Agents.
- 2026-07-08: Logan Kilpatrick announced major Managed Agents updates in the Gemini API, including background task support, remote MCP and function calling, plus network credential refresh; these became available on the free tier.
- 2026-08-09: Harrison Chase highlighted Anthropic’s managed agents engineering page with the summary: “Separate brain from hands,” reinforcing the architecture as a core design pattern.
- 2026-09-04: Philipp Schmid described Managed Agents as a simple way to use Gemini 3.8 Flash in an agentic setup via a single API call that provisions a dedicated remote sandbox with coding tools, network access, Google Search, function calling, MCP, persistent filesystems, background tasks, and cron triggers.
- 2026-11-17: Philipp Schmid recapped Google DeepMind’s self-hosted sandboxes for Managed Agents as one of seven notable developer launches from the prior week.
Relevance to AI PMs
1. Design safer agent products. Managed-agent patterns help PMs define where an agent can act, what tools it can use, and how execution is isolated. This is critical when scoping features involving code execution, browser or network access, enterprise APIs, or file operations.2. Plan product capabilities beyond chat. The mentions show managed agents expanding into background tasks, cron triggers, persistent filesystems, remote MCP, and function calling. PMs can use these as a checklist for turning a conversational assistant into an operational workflow product.
3. Evaluate platform tradeoffs. Managed agents appear in both Anthropic and Google stacks, but with different packaging: architecture guidance, deployable templates, dedicated remote sandboxes, and self-hosted sandbox options. PMs can compare hosted convenience versus infrastructure control based on security, compliance, cost, and time-to-launch.
Related
- Anthropic / anthropic-engineering: Introduced the “brain from the hands” framing and published system-design guidance for scaling managed agents.
- Google DeepMind / Gemini API / Google AI Studio: Positioned Managed Agents as a Gemini-era execution layer with remote sandboxes, free-tier access, and later self-hosted sandbox support.
- Gemini 3.8 Flash / gemini-38-flash: Cited as a model that can be used inside a Managed Agents setup.
- Gemini Decisions API / gemini-decisions-api: Related through Google DeepMind’s broader push toward agentic and decision-oriented developer tooling.
- Background execution / remote MCP / custom function calling: Core capabilities repeatedly associated with managed-agent environments.
- Claude / Claude Code / Cowork: Examples of agent products and deployment surfaces where Managed Agents were used for templates and workflows.
- Harrison Chase, Logan Kilpatrick, Philipp Schmid: Key amplifiers who surfaced architecture ideas, platform updates, and launch recaps.
- Brain from the hands: A shorthand design principle closely tied to managed-agent architecture.
Frequently asked questions
What is managed agents?
A managed-agent infrastructure pattern with self-hosted sandboxes mentioned in Google DeepMind's launches. It reflects the growing emphasis on secure, controllable agent execution environments.
What should AI product managers know about managed agents?
Managed agents separate model reasoning from action execution, improving control, security, and scalability. Anthropic framed the pattern as decoupling the “brain” from the “hands” in large-scale agent systems. Google’s Managed Agents added remote sandboxes, background tasks, MCP, function calling, and later self-hosted sandboxes.
Newsletter Mentions (6)
“#2 𝕏 Philipp Schmid recapped seven Google DeepMind developer launches shipped during the prior week: Gemini 4 Carbon, Gemini 4 Hydrogen, Gemini Decisions API with Gemini Decision 1, self-hosted sandboxes for Managed Agents, a Multimodal Knowledge Agent benchmark, Gemini Robotics 2.5, and AlphaGenome 2. Also covered by: @Philipp Schmid”
#2 𝕏 Philipp Schmid recapped seven Google DeepMind developer launches shipped during the prior week: Gemini 4 Carbon, Gemini 4 Hydrogen, Gemini Decisions API with Gemini Decision 1, self-hosted sandboxes for Managed Agents, a Multimodal Knowledge Agent benchmark, Gemini Robotics 2.5, and AlphaGenome 2. Also covered by: @Philipp Schmid
“Philipp Schmid shared that Managed Agents offers an easy way to try Gemini 3.8 Flash in an agentic environment, providing a dedicated remote sandbox in a single API call.”
Philipp Schmid shared that Managed Agents offers an easy way to try Gemini 3.8 Flash in an agentic environment, providing a dedicated remote sandbox in a single API call. It supports coding tools, full network access, Google Search, function calling, MCP, persistent filesystems, background tasks, and cron triggers through the Google AI Studio free tier.
“#4 𝕏 Harrison Chase shared Anthropic’s “Managed agents” engineering page, writing: “Separate brain from hands.””
GenAI PM Daily August 09, 2026 GenAI PM Daily 🎧 Listen to this brief 3 min listen Today's top 10 insights for PM Builders. Claude Code sessions can now message each other #3 𝕏 Harrison Chase shared a 20-minute explanation of Managed Deep Agents, which he said had launched the previous day and combines a deep agents harness with managed LangSmith infrastructure. He also planned to film more tutorials over the weekend. #4 𝕏 Harrison Chase shared Anthropic’s “Managed agents” engineering page, writing: “Separate brain from hands.”
“Logan Kilpatrick rolled out major updates to Managed Agents in the Gemini API—adding background task support, remote MCP & function calling, and network credential refresh—and now you can try them on the free tier.”
#2 𝕏 Logan Kilpatrick rolled out major updates to Managed Agents in the Gemini API—adding background task support, remote MCP & function calling, and network credential refresh—and now you can try them on the free tier. Also covered by: @Logan Kilpatrick
“Claude launched ready-to-run financial services agent templates for building pitches, conducting valuation reviews, and month-end book closes, available as plugins in Cowork and Claude Code or deployable via cookbooks as Managed Agents.”
#3 𝕏 Claude launched ready-to-run financial services agent templates for building pitches, conducting valuation reviews, and month-end book closes, available as plugins in Cowork and Claude Code or deployable via cookbooks as Managed Agents. Also covered by: @Claude , @Claude Code Blog
“Scaling Managed Agents: Decoupling the brain from the hands - Discusses architecture and design principles for scaling managed agents by separating the decision-making 'brain' from execution 'hands', enabling more robust, scalable agent systems.”
#4 📝 Anthropic Engineering Scaling Managed Agents: Decoupling the brain from the hands - Discusses architecture and design principles for scaling managed agents by separating the decision-making 'brain' from execution 'hands', enabling more robust, scalable agent systems. The article examines tradeoffs and system patterns for large-scale agent management.
Related
An AI company known for Claude. The newsletter references its S-1 and Claude's role in code contributions and design guidance.
Anthropic’s coding agent/product mentioned as a supported client for the ffmpeg-skill repository. It is cited as part of an offline video-editing workflow for agents.
Anthropic's AI assistant, referenced here for its strong role in codebase contributions and design workflows. The newsletter highlights Claude as a practical tool for coding and product design.
AI developer advocate and commentator who recapped Google DeepMind's recent launches. He is cited as a source covering the launches in the newsletter.
Founder/leader associated with LangChain, cited twice in posts about agent architecture and evaluation. He appears in the newsletter sharing and commenting on workflow choices.
Google's AI research organization. It is highlighted for multiple developer launches and for expanding the Google Antigravity program with Gemini access.
AI product and developer relations leader associated with Google. Included among the accounts covering the Genie 4 announcement.
Google’s AI builder environment where Gemini Decisions API was announced. Relevant to PMs as a place to prototype and operationalize Gemini-based workflows.
Google’s API for accessing Gemini models. The newsletter says Gemini 3.7 Flash is available in it and being rolled out to paid users.
Cowork is an Anthropic product mentioned as part of Claude’s product surface. The newsletter references it only as one of the products covered by Anthropic’s containment approach.
Anthropic’s engineering organization, credited here for a detailed post about containing Claude across products. This is relevant to PMs because it addresses agent safety, deployment blast radius, and product containment patterns.
An API announced by Google DeepMind that appears to support decision-making workflows for agents and developers. It is presented alongside a related model, Gemini Decision 1.
Stay updated on managed agents
Get curated AI PM insights delivered daily — covering this and 1,000+ other sources.
Subscribe Free