What is an agent gateway?
An agent gateway is a network proxy on the connections between AI agents and the tools, agents and models they call, authenticating each party and enforcing access and content policies in one place.
Checked
As a company runs more agents, each with its own MCP servers, APIs and peer agents, security policy ends up scattered across codebases. An agent gateway pulls it into the network path: every call goes through a proxy that knows which agent is calling, what it may reach and what the traffic contains.
Google announced Agent Gateway at Cloud Next '26 in April 2026 as part of the governance stack in its Gemini Enterprise Agent Platform. Amazon Bedrock AgentCore Gateway covers overlapping ground, with more emphasis on turning existing APIs into agent tools.
What Google's Agent Gateway does
Ingress and egress
Client-to-Agent mode sits in front of hosted agents or MCP tools and authenticates callers. Agent-to-Anywhere mode governs outbound calls from agents to tools, MCP servers, SaaS APIs and other agents.
Protocol aware
It reads MCP JSON-RPC requests to see tool names and methods, governs A2A delegation between agents, and handles REST and gRPC traffic.
Deny by default
IAM policies link agent identities to approved tools and endpoints, and connections are blocked unless a policy grants them. The gateway checks Agent Registry, the directory of approved agents, tools and MCP servers.
Content inspection
Model Armor scans prompts and tool responses for prompt injection, sensitive data leaks and harmful content.
Availability
Google's documentation says it currently governs agents on Agent Runtime and Gemini Enterprise, with GKE, Cloud Run and cross-cloud support coming through early preview, and a limit of 5,000 registered resources per gateway instance.
Same name, different emphasis
AWS describes Amazon Bedrock AgentCore Gateway as a managed entry point for agent traffic that converts APIs, Lambda functions and existing services into MCP-compatible tools, fronts other agents, routes model calls across providers and handles both inbound and outbound authentication, with semantic search over available tools. Google's gateway starts from governing traffic; AWS's starts from making tools available. Both authenticate in both directions.
An agent gateway also differs from an LLM or AI gateway, which proxies an application's calls to model APIs for routing, metering and caching. Some products, AgentCore Gateway among them, now do both jobs.
What a gateway cannot see
Local actions
Code that runs inside the agent's sandbox, file writes and local shell commands never cross the gateway. The sandbox governs those.
Intent
A gateway sees a permitted tool called with permitted arguments. It cannot tell whether the user wanted that call; approval and classifier gates closer to the agent handle that.
Novel attacks
Content scanning is classifier-based and will miss some injections. Deny-by-default access policy is the deterministic part, and the part to rely on.
How Motiontic's engineers apply it
Motiontic's engineers route agent tool traffic through the client's existing gateway where one exists. Where none does, egress allowlists and credential injection sit in a proxy outside the agent, which does the same job on a smaller scale. Policies live in version control next to the agent's tool definitions, so a new tool cannot ship without its access rule.
Agent gateway compared with related controls
| Agent gateway | AI (LLM) gateway | Agent sandbox | |
|---|---|---|---|
| Sits between | Agents and the tools, agents and models they call | Applications and model APIs | Agent code and the host system |
| Main job | Authenticate, authorise and inspect agent traffic | Route, meter and cache model calls | Limit file system and network reach |
| Sees | Every network call the agent makes | Prompts and completions | Everything the agent does locally |
| Misses | Local actions and user intent | Tool calls to other systems | Misuse of allowed channels |
Common questions
What is an agent gateway?
An agent gateway is a proxy that every AI agent connection passes through, so one place can authenticate agents, decide which tools and agents they may reach, and inspect the traffic for threats such as prompt injection.
How is an agent gateway different from an API gateway?
An API gateway fronts services for any caller. An agent gateway understands agent protocols such as MCP and A2A, ties policy to agent identities and can inspect tool names and prompt content.
Is Google's Agent Gateway available now?
Google's documentation says it currently works with agents on Agent Runtime and Gemini Enterprise, with GKE, Cloud Run and cross-cloud support coming through early preview. Some partner integrations announced at Next '26 were described as coming to preview.
Does an agent gateway stop prompt injection?
It reduces the risk by scanning content and blocking unapproved connections, but it does not see what happens inside the agent. It works alongside a sandbox, least privilege and approvals, not instead of them.
Related
Agent identity
Each agent gets its own credential, permissions and audit trail.
Model Context Protocol (MCP)
The open standard for plugging tools and data into AI agents.
A2A protocol
The open protocol for agents talking to other agents, now under the Linux Foundation.
Agent sandbox
Hard walls around where an agent can read, write and connect.
AI transformation
Moving real operations onto AI: the data layer, the workflows and the proof that it paid off.
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.