共有:

Agent Runtime · Gemini API

The day the HTTP keep-alive
disappears from long-running agents.

Google added background execution to the Gemini API, plus native support for connecting to remote MCP servers. The real move isn't async endpoints — it's that where the agent lives shifts from client to server.

AI Navigate Editorial·2026.07.19·6 min read

HTTP KEEP-ALIVE Client held open the whole run Server disconnect = fail BACKGROUND EXEC Client Gemini API (queue) Remote MCP tools fire & forget
FIG. The old model tied the agent's lifetime to a client-held HTTP connection. The new one lifts that responsibility onto the server.
01

What Was Announced

Gemini API grows a "does-not-die" agent

Per the Gemini API documentation, the API now supports background execution — long-running tasks that survive without a client-held HTTP connection — plus connection to remote MCP (Model Context Protocol) servers. MCP is the open spec originally proposed by Anthropic, and Google adopting it materially widens tool-connection interop.

Two things are worth naming. First, the assumption "long-lived agent = permanently open HTTP" is gone. Second, tool connections are converging on MCP rather than each vendor's private protocol. The first is an operations shift; the second is an ecosystem shift.

02

By The Numbers

The architectural core in three moves

0 conn
no client-held connection
MCP
shared protocol widens tool interop
Async
Job-queue mental model becomes default
03

Why It Matters Now

Why the timing matters

HTTP-held implementations have quietly been the top bug source for long-running agents this year.

Coding and research agents now routinely run 30 minutes to hours. As lifetimes grew, load-balancer disconnects, timeouts, and NAT resets climbed to the top of the failure list. WebSocket or SSE hacks help but don't fix the fundamental: the moment the pipe breaks, the agent is stranded. The Gemini API change moves lifetime management from client to server, and the disconnect risk simply leaves the diagram.

MCP support isn't just "one more integration." ChatGPT with GPTs, Claude Tool Use, Gemini Function Calling — each demanded its own connector code. Under MCP, the same tool server can be called by all of them. Anthropic, OpenAI, Cursor, and Zed are already on it; Google joining accelerates the "de facto standard" trajectory visibly.


04

Who Should Do What

Who this hits, and how to act

The impact is heavy for long-lived-agent teams and near-zero for one-shot API callers.

01

Long-lived agent devs: go serverless again

"Keep an EC2 instance alive to hold the HTTP" has been the default pattern. Now a short-lived process on Cloud Run / Lambda can submit and poll by job id later, letting many teams delete their long-running VM tier from Terraform.

02

Tool providers: MCP-server your API

SaaS vendors that wrap their APIs behind a thin MCP server open a single entry point that Claude, ChatGPT, Gemini, and Cursor can all hit equally. "MCP-compatible" is climbing procurement checklists faster than "plugin for X."

03

One-shot callers: no real change

Single-prompt text generation or embedding is untouched. A clean rule of thumb: enable background execution only when a task's expected wall time exceeds ~30 seconds.

05

What Comes Next

What to expect, what to do

Job API
submit/poll pattern becomes default
MCP
registry & auth-broker services emerge
Cost
idle keep-alive cost recovered

Near-term outlook. (a) OpenAI Responses API and Anthropic Messages Batch are already moving the same way — with Google joining, "LLMs as async job queues" solidifies as a shared abstraction. (b) Expect an MCP registry and permission-brokering middleware. (c) The concrete first move for teams is a design review of any daemon-style agent — pulling stats on failures caused by HTTP disconnects makes the ROI of the migration easy to size.


Running an agent for hours is,
at bottom, a question of who owns its lifetime.


06

Counterpoint

The other side of the argument

First, async is harder to debug. A single request-per-session model kept logs linear; jobs make mid-run state something you have to actively fetch, and reproducing why it stalled gets more expensive. Observability tooling — traces, spans — is no longer optional.

MCP itself isn't a clean win either. Permission scopes and sandboxing remain looser than production comfort — reaching a remote MCP server means the agent trusts what it returns for the next tool call. Adversarial MCP servers become a fresh prompt-injection surface, and that risk needs to be modeled from day one, not as a follow-up.