Aakib Ansari.
Back to articles
News Brief

Gemini's Managed Agents API Now Runs Jobs in the Background and Connects Directly to Remote MCP Servers

Md Aakib Ansari
Md Aakib AnsariWeb Developer & AI Tools Reviewer
3 min readModel: Gemini 3.6 Flash
Gemini's Managed Agents API Now Runs Jobs in the Background and Connects Directly to Remote MCP Servers

The core problem with running long agent tasks over a standard HTTP connection is fragility: if anything breaks the connection — a network blip, a timeout, a client restart — the job dies mid-flight. Google's solution, shipped this month to the Managed Agents feature in the Gemini Interactions API, is a straightforward flag: set background: true in your interaction request, and the API hands back an interaction_id immediately instead of holding the connection open. Your client can walk away, reconnect hours later using that ID, and retrieve the result — or stream progress updates while it runs in Google's isolated cloud sandbox. For agents doing anything that takes more than a minute or two, this changes the reliability calculus entirely.

The second update shipped alongside it: Managed Agents now connect directly to remote Model Context Protocol (MCP) servers without requiring developers to build custom proxy middleware. Pass an mcp_server tool in the interaction request and the agent can reach private databases, internal APIs, or any other resource sitting behind an MCP endpoint — mixing it with built-in sandbox tools like Google Search or code execution in the same call. The infrastructure side — HTTP endpoints, authorization, audit logging — is handled on Google's end rather than requiring a local stdio server.

Both updates land in the Gemini Interactions API, which went generally available in June 2026 and is now the recommended interface for all new Gemini agent work. Gemini 3.6 Flashwhich we covered at its release in July — is the default model powering Managed Agents. Enterprise Agent Platform users also got an extended runtime cap this month: up to seven days of continuous execution for complex multi-step workflows, up from the shorter windows available before.

A credential refresh mechanism also shipped at the same time, maintaining authentication across long-running sessions without manual re-authentication — a detail that matters for any agent expected to stay operational over hours rather than minutes.

Frequently Asked Questions

What does the background: true flag actually do?
It switches the Gemini Interactions API from a synchronous, connection-held response to an async one. The API immediately returns an interaction_id; your client uses that ID to poll status, stream progress, or retrieve the final result after the agent finishes — regardless of whether the original connection stayed open.
Does remote MCP support replace local MCP servers?
It's a complement, not a replacement. Local MCP servers using stdio still work. Remote MCP runs on Google's infrastructure with managed HTTP endpoints, centralized discovery, and audit logging — better suited for production deployments connecting to private internal services.
Which model powers Managed Agents?
Gemini 3.6 Flash is now the default. You can specify a different model in the interaction request if needed.

Related Articles

Meta Releases Muse Glimmer: Apache 2.0 Licensed 30B Local Agent Model
News Brief2 min read
Meta Releases Muse Glimmer: Apache 2.0 Licensed 30B Local Agent Model

Meta Superintelligence Labs has released Muse Glimmer, a 30-billion parameter open-weight model distilled from its proprietary Muse Spark flagship. Published under an Apache 2.0 license, Glimmer is purpose-built for offline, on-device agentic workloads like coding, debugging, and file management on consumer hardware.

Ant Group's inclusionAI Team Releases Ling 3.0 Flash FP8 Under MIT License
News Brief2 min read
Ant Group's inclusionAI Team Releases Ling 3.0 Flash FP8 Under MIT License

Ant Group's inclusionAI team has released Ling 3.0 Flash FP8, a highly efficient 124-billion parameter Mixture-of-Experts (MoE) model. Featuring an MIT license and a custom hybrid attention architecture, the model reduces active parameters to 5.1 billion per token, matching the performance of much larger models while dramatically lowering operational costs.

Liquid AI's LFM2.5-2.6B Matches Models Three Times Its Size on Agentic Tasks
News Brief2 min read
Liquid AI's LFM2.5-2.6B Matches Models Three Times Its Size on Agentic Tasks

Liquid AI released LFM2.5-2.6B on August 4, a 2.69B-parameter on-device model purpose-built for agentic workloads. Using a hybrid architecture of short convolution blocks and grouped query attention, it runs under 2.5 GB of memory and reaches approximately 220 tokens/s on Apple M5 Max — while matching or exceeding Qwen3.5-9B on tool use and instruction-following benchmarks.