Vertex AI became Gemini Enterprise Agent Platform. Model retirements, planned ahead.
Google folded Vertex AI into Gemini Enterprise Agent Platform in April 2026, and Gemini 2.5 Pro, Flash and Flash-Lite retire on 20 October 2026. We move your workloads onto current Gemini models. New agents we build with ADK on Agent Runtime.
Vertex AI becomes Gemini Enterprise Agent Platform22 Apr 2026
Gemini 2.5 models retire20 Oct 2026
connects with
in short·15-second read
Cohash works on Gemini Enterprise Agent Platform, which is what Google has called Vertex AI since April 2026.
We move applications off retiring models like Gemini 2.5. We also build agents with the Agent Development Kit, run them on Agent Runtime (Agent Engine, before the rename) and connect them to BigQuery and business systems through scoped service accounts. Every build ships with an evaluation set.
you probably need this if
Something in production still calls a Gemini 2.5 model, and we're at the retirement date.
Agents run on Agent Runtime under a default service account with broad access.
Nobody can say which documents the search index can return to which users.
22 Apr 2026Vertex AI becomes Gemini Enterprise Agent Platform Google's line is that this is the evolution of Vertex AI. On the ground, Agent Engine is now Agent Runtime, and new Vertex AI services come out through the Agent Platform. source ↗
20 Oct 2026Gemini 2.5 models retire Gemini 2.5 Pro, Gemini 2.5 Flash and Gemini 2.5 Flash-Lite retire on 20 October 2026. Gemini 2.0 Flash retired on 1 June 2026. source ↗
today · Oct 22, 2026
no date yetModel retirement noticeGoogle posts fixed retirement dates with at least 45 days to migrate, and it won't bring them forward.source ↗
the story · gemini enterprise agent platform (vertex ai) estate
From the first call to a setup your own team can run.
Here's what usually prompts a call about your Gemini Enterprise Agent Platform (Vertex AI) estate, what we do about it and in which order, and where you end up.
01 · where it hurts02 · what we change03 · how it runs04 · where you end up
incident board · gemini enterprise agent platform…3 open
openSomething in production still calls a Gemini 2.5 model, and we're at the retirement date.
recurringAgents run on Agent Runtime under a default service account with broad access.
blockedNobody can say which documents the search index can return to which users.
exposure
Gemini model migrationsAgents with ADK and Agent RuntimeRetrieval and Agent SearchIdentity, networking and data…Evaluations and monitoringGemini Enterprise rollout
startrehearsehandover
Model and agent inventory
Migration and build plan
Evaluate and switch
Handover
we are here
ready to hand over
Apps go across to a supported model in a known order.
The agent is approved by security on the evidence.
Model version register with retirement dates and owner per application
Gemini migration plan with evaluation results per application
chapter 01 · where it hurts
Why teams reach out
Something in production still calls a Gemini 2.5 model, and we're at the retirement date.
Agents run on Agent Runtime under a default service account with broad access.
Nobody can say which documents the search index can return to which users.
chapter 02 · what we change
How we solve it
Gemini model migrations
Agents with ADK and Agent Runtime
Retrieval and Agent Search
Identity, networking and data access
Evaluations and monitoring
Gemini Enterprise rollout
chapter 03 · how it runs
Our approach
Model and agent inventory
Migration and build plan
Evaluate and switch
Handover
chapter 04 · where you end up
What changes
Apps go across to a supported model in a known order.
The agent is approved by security on the evidence.
Model version register with retirement dates and owner per application
Gemini migration plan with evaluation results per application
what we do
Our Gemini Enterprise Agent Platform (Vertex AI) work
6 kinds of Gemini Enterprise Agent Platform (Vertex AI) work, each run by engineers who have done it on live tenants.
Applications moved off retiring Gemini versions to current ones, with prompts retested on an evaluation set and the switch made behind a configuration flag.
Agents written with the open-source Agent Development Kit in Python, TypeScript, Go or Java, deployed to Agent Runtime, Cloud Run or GKE depending on your platform.
Grounding on your documents and data with Agent Search, formerly Vertex AI Search, with sources, access rules and metadata reviewed before indexing.
Service accounts, Workload Identity Federation, VPC Service Controls and IAM roles designed so agents reach only the data and tools they need.
Evaluation sets from real cases, run on every prompt or model change, with logging and cost alerts in Cloud Monitoring.
Agents published to employees through the Gemini Enterprise app, with groups and data connectors configured per team.
capability 01 / 06in production
Gemini model migrations
Applications moved off retiring Gemini versions to current ones, with prompts retested on an evaluation set and the switch made behind a configuration flag.
Agents written with the open-source Agent Development Kit in Python, TypeScript, Go or Java, deployed to Agent Runtime, Cloud Run or GKE depending on your platform.
Gemini Enterprise Agent Platform (Vertex AI) situations we walk into
We pieced these together from Gemini Enterprise Agent Platform (Vertex AI) situations that come up a lot. They're patterns. You won't find a specific client in them.
case file 01
Eleven apps still on Gemini 2.5 Flash on the day it retires
composite · not a client story
Beforethe situation
Eleven internal apps call Gemini 2.5 Flash through Vertex AI endpoints, and that model retires on 20 October 2026. Just two have any kind of test set.
What we dothe work
We pull every caller out of the logs and rank them by business impact. The top apps get small evaluation sets written from real requests. Then, highest impact first, each one moves to a current Gemini Flash model behind a configuration flag.
Afterwhere you end up
Apps go across to a supported model in a known order. Every app keeps a test set, ready for the next migration.
case file 02
Security won't approve an Agent Engine prototype
composite · not a client story
Beforethe situation
Someone built a procurement agent with ADK and it runs on Agent Runtime, formerly Agent Engine, under the default compute service account. So it can read every BigQuery dataset in the project.
What we dothe work
We give the agent its own service account with read access to two views and put the project inside VPC Service Controls. Every tool call gets logged. An evaluation set runs before a staged release.
Afterwhere you end up
The agent is approved by security on the evidence. Its access is whatever the procurement process actually needs.
case studies
Gemini Enterprise Agent Platform (Vertex AI) case studies
How a Gemini Enterprise Agent Platform (Vertex AI) engagement runs
We don't start a step until the last one has produced something you can check.
step 01
Model and agent inventory
We list every application, agent and pipeline calling Gemini or other models on the platform, with model version, retirement date, owner and service account.
step 02
Migration and build plan
Retiring models get replaced first, since those have a date on them. Then agents are built or hardened with ADK, Agent Runtime and scoped IAM, in whatever order the retirement dates and business risk put them.
step 03
Evaluate and switch
We score every change on an evaluation set made from real cases. It's switched over behind a flag, and logging and cost alerts stay on.
step 04
Handover
We hand your team a model version register and the evaluation pipelines, along with runbooks for when the next retirement notice comes.
ai on gemini enterprise agent platform (vertex ai)
Where AI fits on Gemini Enterprise Agent Platform (Vertex AI), and where it needs guardrails.
Vertex AI is now Gemini Enterprise Agent Platform, announced on 22 April 2026. Agents built with the Agent Development Kit run on Agent Runtime, which was Agent Engine. Google also offers managed remote MCP servers for BigQuery and the Agent Platform. Gemini 2.5 models retire today, 20 October 2026, so for us model migration comes first.
AI agentasks for data or an action
Scoped identityonly the access it needs
MCP serverthe one door into the platform
Logged, approveda person signs off on writes
Gemini Enterprise Agent Platform (Vertex AI)does the work
There are managed remote MCP servers from Google, for BigQuery and Gemini Enterprise Agent Platform among others. They implement the MCP authorization spec, and access per resource is set by IAM allow and deny policies. We give each agent its own identity and start it on read-only roles. Writes are logged and need approval.
Gemini model migration
Gemini 2.5 Pro, Flash and Flash-Lite retire on 20 October 2026. For us that means finding every caller in the logs, testing current Gemini models on an evaluation set and switching behind a configuration flag.
ADK and Agent Runtime
The open-source Agent Development Kit builds agents in Python, TypeScript, Go or Java, deployed to Agent Runtime, Cloud Run or GKE. For us the rule is one service account per agent, with narrow IAM roles.
Agent Search for grounding
Agent Search (it used to be Vertex AI Search) grounds answers on your documents. For us the work is reviewing sources and access rules before indexing, so an agent only returns content its user may see.
Managed MCP servers
Google runs remote MCP servers for its own services. BigQuery's lives at bigquery.googleapis.com/mcp, and the Agent Platform has one too. We use them instead of self-hosting wherever they fit.
How agents on Gemini Enterprise Agent Platform reach users and data
Agents built with ADK run on Agent Runtime and call Gemini models through the Agent Platform. They ground answers with Agent Search and query BigQuery or Snowflake under scoped service accounts. Employees get to them from the Gemini Enterprise app or from a business app.
What a Gemini Enterprise Agent Platform (Vertex AI) engagement leaves behind
These are working documents, written against your configuration, so your team can run them without us.
inventory01
Model version register with retirement dates and owner per application
yours to keep
plan02
Gemini migration plan with evaluation results per application
yours to keep
code03
ADK agent code deployed to Agent Runtime with CI
yours to keep
design04
Service account and IAM design per agent
yours to keep
inventory05
Agent Search source inventory with access rules
yours to keep
runbook06
Runbooks for model switches and incident response
yours to keep
talent and delivery · gemini enterprise agent platform (vertex ai)
Certified engineers, wherever you need them.
Some clients want a project team, some want a couple of engineers on theirs. Either works. Onsite, nearshore or offshore, with handovers across time zones that keep things moving.
Forward deployed AI engineer (Google Cloud, Gemini)
ADK agent developer (Python, TypeScript)
Google Cloud AI platform engineer (IAM, VPC-SC)
ML engineer for Gemini model migration and tuning
LLM evaluation engineer
due diligence · gemini enterprise agent platform (vertex ai)
Questions to ask any partner, including us.
Put these to anyone bidding on your Gemini Enterprise Agent Platform (Vertex AI) work. The ones who've done it answer differently from the ones who've read about it.
Q1
Which apps of ours still call a Gemini version with a retirement date in the next six months?
Google gives at least 45 days' notice, but large estates need longer to test. The inventory should exist before the notice arrives.
Q2
What service account will each agent run as, and what can it read?
Default service accounts often have broad project access. Security approval starts with an account per agent and narrow roles on it.
Q3
How will you test prompt and model changes before production?
No two Gemini versions behave quite the same. Upgrade without an evaluation set built from real cases and you're guessing.
Q4
Which documents can Agent Search return, and to whom?
Indexing a shared drive can expose content to people who shouldn't see it. We review the sources and set the access rules first. Then we index.
Q5
How do the old Vertex AI names map to what we are licensing now?
Product names changed in April 2026. Old names may still be in contracts, IAM roles and documentation, and the team needs a clear map between them.
01Is Vertex AI still a product?
It lives on as Gemini Enterprise Agent Platform, which Google announced on 22 April 2026 as the evolution of Vertex AI. Vertex AI services and roadmap items ship through the Agent Platform. Several components were renamed along the way, Agent Engine to Agent Runtime among them.
02Our app calls Gemini 2.5. What do we do now?
You have until 20 October 2026. That's the retirement date Google's model versions page gives for Gemini 2.5 Pro, Flash and Flash-Lite. Move to a current Gemini model now, and run your evaluation set against it before production traffic switches.
03How much notice does Google give before retiring a model?
Google says retirement dates give at least 45 days to migrate and won't be moved earlier. We keep a register of model versions per application so the next date is planned, not discovered.
04Should we build agents with ADK or another framework?
ADK is Google's own open-source framework. It deploys straight to Agent Runtime, Cloud Run or GKE. If your team already has another framework, that will still run on Google Cloud. Which one we pick comes down to your languages, your hosting and the tools the agent needs.
05Do you embed engineers with our business team?
Yes. Our forward deployed engineers sit with your team and build against your real systems and data. When they hand over, you get the evaluations, access controls and runbooks. We staff onsite in the US, or nearshore and offshore, in all time zones.
services
Related services
Projects on Gemini Enterprise Agent Platform (Vertex AI) usually end up touching one of these, so they're worth a look.
you are hereGemini Enterprise Agent Platform (Vertex AI)
Reviewed by the Cohash teamGemini Enterprise Agent Platform (Vertex AI) is a trademark of its owner, used here to identify the platform.how we handle facts and data →
next step
Tell us what you're running. We'll tell you what we'd do first.
A 30-minute working session with a senior engineer. No slides.
