What is an AI agent platform?

An AI agent platform is the software layer where agents are built, deployed, governed, and run. It owns the agent loop, handles credentials and permissions, stores durable state, connects to systems of record, and keeps an audit trail. Judge one by what survives after a run ends. On Major, agents build deterministic apps for repeatable work, so the lasting artifact is working software with its own database and logs.

Is Gemini Enterprise Agent Platform the same as Vertex AI?

It is Google's renamed and expanded evolution of Vertex AI, announced 22 April 2026, and Google's product page still says formerly Vertex AI. It adds agent integration, DevOps, orchestration, and security across four pillars: Build, Scale, Govern, and Optimize. Its agents reason each time they execute, whereas on Major an agent writes an app once and the app handles the repeatable work.

What is the difference between an agent platform and an agent framework?

A framework decides what the agent does next, giving you reasoning loops, state transitions, and handoffs. A platform supplies everything around it: runtime, triggers, persistent state, permissions, integrations, and activity records. Every framework needs a platform beneath it. Major sits at the platform layer and adds an app layer, so work an agent turns into code stops depending on the reasoning loop.

Is an agent platform the same as an LLM gateway?

No. An LLM gateway sits in front of the models, routing requests, capping spend, and logging prompts. Control ends when the agent acts on a business system. An agent platform owns the action: the scoped credential, the record written, and the attributable audit entry. Teams often run both. On Major, repeatable work becomes an app whose database and logs are the durable record.

What should I look for in an agent platform?

Start with where state lives after a run ends: where is the output thirty days later and who can query it. Then check how credentials are scoped and whether actions produce attributable audit records. Ask which steps run as code instead of reasoning, which open protocols it speaks, how you debug a past run, and what exports if you leave.

Agent Platform: What It Means Now That Google Owns the Term

Google renamed Vertex AI to Gemini Enterprise Agent Platform and now owns most of this search result. Here is the neutral definition, the criteria that decide the choice, and how to build on Major.

Jason BaoAugust 19, 2026Updated October 22, 2026

Platform team gathered around a whiteboard planning how agents will be built and governed

An AI agent platform is the software layer where agents run in production. It owns the agent loop, the credentials an agent acts under, the state it leaves behind, the connections to your systems, and the audit record of what it did. If you are a platform or IT leader choosing one, judge it by a single question: what exists thirty days after a run ends?

Major answers that question with software. On Major, an agent that works out a repeatable part of a job builds an app for it, with its own managed database, file storage, and logs, then runs the app instead of reasoning through the step again. Reason once, run forever. The thing you govern is code you can read, not a transcript.

Why "agent platform" now names a product too

On 22 April 2026 Google announced Gemini Enterprise Agent Platform, which it describes as "the evolution of Vertex AI" with new features for agent integration, DevOps, orchestration, and security. Its product page still carries "formerly Vertex AI" in the title. Google organizes it into four pillars: Build, Scale, Govern, and Optimize.

So half the people typing "agent platform" want Google's product page, and the other half want to know what the category requires. A vendor announcement cannot answer the second question honestly, because the answer would be its own feature list. Fix your own definition first, then hold every product against it.

The portable version: an agent platform is the software layer where agents are built, deployed, governed, and run, and it should be judged by what survives after a run ends.

Platform, framework, gateway, builder: what each one covers

A framework decides what the agent does next. LangGraph, CrewAI, the OpenAI Agents SDK, and Microsoft Agent Framework live here. They leave hosting, scoped credentials, durable business data, access control, and audit to your team.

A platform wraps execution infrastructure around that loop: runtime, triggers, persistent state, permissions, system connections, and an action record. Every framework needs a platform underneath it.

An LLM gateway sits in front of the models. It routes requests, caps spend, and logs prompts. That is control over model traffic and none over what the agent then does in Salesforce or Postgres.

An agent builder is a visual or natural-language design surface. Without durable runtime state, deployment, permissions, and audit, it produces a prototype.

Checkpoints and session memory preserve reasoning context. They do not preserve the business object the work was about. That gap is where most agent projects stall.

The seven criteria that decide the choice

Work through these in order against any shortlist. The first two predict production survival better than the rest.

Where state lives after the run ends. Ask where the output exists next month, who can query it, and whether it is more than chat history.
Credentials, permissions, and audit. A trace says the model called a tool. An audit record says which identity changed which row and who authorized it.
What the model does versus what runs as code. Steps the model repeats cost tokens every time and vary every time. Steps that run as code do neither.
Who owns the agent loop. Can your team read and change the control flow, or does the platform infer it from a prompt?
Protocols. Model Context Protocol lets a tool you build once serve every agent you run. A platform that reaches your systems only through its own connector catalog sets your integration roadmap for you.
Observability. Take a run from last month that went wrong. Can you explain it, reproduce it, and fix the cause? Replaying a conversation is not reproducing a failure.
Exit. Export an agent and look at what you get. Code and data travel. Proprietary flow definitions do not.

How the main approaches answer those questions

Five approaches compete for this work.

Major: agents build deterministic apps for repeatable work, so state sits in a managed database, credentials are scoped at the point of action, and every action is logged in code the team can inspect. Built for internal work that several people depend on and auditors will ask about.
Hyperscaler agent platforms (Google Gemini Enterprise Agent Platform, plus AWS and Microsoft equivalents): runtime, managed memory, and platform identity controls, with agents that reason each time they execute. Their strength is model access, training, and MLOps at cloud breadth.
Developer frameworks (LangGraph, CrewAI, OpenAI Agents SDK): explicit control over transitions, with checkpoints for state. Hosting, durable data, and audit are yours to build.
Conversational and CX platforms (Sierra, Cognigy, Rasa, Agentforce, Copilot Studio): dialogue-shaped agents with state held in conversations and the surrounding CRM. Aimed at customer-facing service.
Automation tools used for agent work (n8n, Zapier, Make): a workflow engine with the model handling one decision step. Run history and connector-level credentials are the governance surface.

The pattern across the last four is the same. Each governs the agent, wrapping permissions and logging around a model that re-reasons the task on every run. You end up holding an audit trail of a black box. Major changes what you hold. The logic the agent figured out is saved as an app, so you can read it, version it, and test it.

Build a governed agent platform on Major

Here is what a platform team stands up on Major for a recurring internal job, using accounts payable review as the example.

Connect the systems. Add the connectors the work touches, for example NetSuite for bills, Gmail for vendor email, and Slack for approvals. Credentials are scoped per connector and applied through the credential proxy, so the agent never holds a raw secret.
Let the agent build the app. The agent works out the repeatable steps (match invoice to purchase order, check terms, flag duplicates) and builds an app with managed database tables for invoices, decisions, and exceptions.
Reserve the model for judgment. The agent reasons only on the ambiguous cases, such as a partial shipment or an odd vendor email. Everything else runs as deterministic code, so token spend is front-loaded into building the app and then stays flat as volume grows.
Trigger it on a schedule or event. A workflow runs the app nightly or when a new bill lands, and routes exceptions to a person in Slack for approval.
Govern it from day one. Role-based access decides who can run, edit, or view the app. SSO, audit logging, and storage come from the platform, and every action is attributable to an identity.
Reuse it. The next team that needs invoice matching runs the same app instead of rebuilding the logic.

Build your first governed agent on Major and run the thirty-day state test yourself.

What an agent platform costs, and what drives it

Price models differ across approaches: frameworks cost staffing, hyperscalers meter tokens, runtime, and storage, conversational platforms meter conversations, and automation tools meter runs. Work out which meter your workflow will hit before comparing anything.

The variable that moves the bill most is how much reasoning happens per run. An agent that re-reasons the same workflow every execution pays for it every execution, and the curve follows volume. Cheaper model routing trims the slope. Moving a step out of the model into an app is the only change that alters the shape of the curve, which is why Major treats it as the default. Cost here is structural: front-loaded, then flat. The model still does the judgment work. It just does less of everything else.

How to choose: a short sequence

Write down one real workflow: trigger, operations in order, systems touched, how it fails.
Mark every step that should behave identically each time. Those are candidates for code.
Run the thirty-day state test. Where does the output live next month, and can a colleague pick it up?
Demand one controlled action in a demo: a scoped credential, an attributable identity, a readable audit record.
Estimate volume and exit cost together.

Steps three and four eliminate most shortlists. Major clears both because the output is an app with its own database and logs, and every action runs under a scoped identity.

Start building on Major

Build the apps and agents behind your own workflow on Major, with the database, permissions, and audit already in place, and see what remains after the run ends.

Frequently asked questions

What is an AI agent platform?: An AI agent platform is the software layer where agents are built, deployed, governed, and run. It owns the agent loop, handles credentials and permissions, stores durable state, connects to systems of record, and keeps an audit trail. Judge one by what survives after a run ends. On Major, agents build deterministic apps for repeatable work, so the lasting artifact is working software with its own database and logs.
Is Gemini Enterprise Agent Platform the same as Vertex AI?: It is Google's renamed and expanded evolution of Vertex AI, announced 22 April 2026, and Google's product page still says formerly Vertex AI. It adds agent integration, DevOps, orchestration, and security across four pillars: Build, Scale, Govern, and Optimize. Its agents reason each time they execute, whereas on Major an agent writes an app once and the app handles the repeatable work.
What is the difference between an agent platform and an agent framework?: A framework decides what the agent does next, giving you reasoning loops, state transitions, and handoffs. A platform supplies everything around it: runtime, triggers, persistent state, permissions, integrations, and activity records. Every framework needs a platform beneath it. Major sits at the platform layer and adds an app layer, so work an agent turns into code stops depending on the reasoning loop.
Is an agent platform the same as an LLM gateway?: No. An LLM gateway sits in front of the models, routing requests, capping spend, and logging prompts. Control ends when the agent acts on a business system. An agent platform owns the action: the scoped credential, the record written, and the attributable audit entry. Teams often run both. On Major, repeatable work becomes an app whose database and logs are the durable record.
What should I look for in an agent platform?: Start with where state lives after a run ends: where is the output thirty days later and who can query it. Then check how credentials are scoped and whether actions produce attributable audit records. Ask which steps run as code instead of reasoning, which open protocols it speaks, how you debug a past run, and what exports if you leave.