On used AI agents to migrate 24 services in two weeks each, not three months
On and Google Cloud say a multi-agent system cut each service migration from three months to two weeks, with under five minutes of planned downtime and humans approving cutovers.
On and Google Cloud announced on 8 October 2026 that On used AI agents to migrate 24 core services to Google Cloud. The release says each migration took two weeks instead of three months, and that a small team of On's own engineers ran 15 of the migrations entirely in-house. It is vendor-published, issued by Google Cloud with a quote from On's chief technology officer, Narek Verdian.
How the work was split
According to the release, the multi-agent system divided each migration into specialised steps: codebase analysis, infrastructure configuration generation, translation pipeline creation and testing validation. Agents were given defined tasks with clear checks. The release says this went beyond standard developer assistants, towards coordinated, domain-specific agents handling migration tasks end to end.
On says its engineers used agentic workflows to map inconsistent deployment patterns across more than 20 code repositories in hours instead of days. Mechanical work such as configuration scripting, data replication and environment validation went to the agents, which On says protected its product engineering roadmap and removed the need for outside implementation help.
Where humans stayed in the loop
The release is specific about control. Engineers reviewed generated code, ran the infrastructure changes and authorised production cutovers, and consequential changes remained subject to human approval. Verdian is quoted as saying human engineers remain accountable for the decisions that matter. On reports under five minutes of planned production downtime per service.
Reading this at work? Get one story like it explained every morning.
A five-minute read every morning at 7am ET: one story explained, two in brief, what else happened, one number. Unsubscribe in one click.
The cost claim and what is missing
The only money figure is a comparison: the release says an external vendor quoted $500,000 to migrate just a fraction of the microservices by traditional manual methods, and that On's team completed the broader project without outside help.
- No cost for the agents, models or tokens is given, so the $500,000 comparison is one-sided.
- The release does not name the models or agent framework used, beyond saying a single agentic platform and secure models were involved.
- It gives no count of defects, rollbacks or review time for the generated code, and no size for the internal team.
- The 24 services and the 15 in-house migrations are On's figures; the release does not say how the other nine were run.
- Gemini Enterprise appears separately: On says it has rolled it out to every employee for no-code agents, which is not the migration tooling.
What a team can take from it
The pattern is a bounded one: break a repetitive engineering project into steps with checkable outputs, let agents do the mechanical steps, and keep a named human on every irreversible action. If you are weighing an outside integrator for a migration, ask them to price the same project with agents and to show review time and defect counts, which this announcement does not.
Questions this article answers
How did On use AI agents for its Google Cloud migration?
On says a multi-agent system split each migration into codebase analysis, infrastructure configuration, translation pipeline creation and test validation, while engineers reviewed code and authorised production cutovers.
How long did On's AI-agent migration take?
On and Google Cloud say each service took two weeks instead of three months, across 24 core services, with under five minutes of planned downtime per service.
Did On publish what the AI agents cost?
No. The 8 October 2026 release cites a $500,000 external vendor quote for a fraction of the services but gives no agent, model or token costs.
Anthropic Cyber Mission: free OSS Scanner and an infrastructure defence program
Anthropic launched its Cyber Mission on 8 Oct: free model scans for open-source projects, sent without human review, plus a program with 11 OT security partners.
Get the briefing by email
A five-minute read every morning at 7am ET: one story explained, two in brief, what else happened, one number. Unsubscribe in one click.
If you have run a migration with agents, which steps did you let them do alone, and which cutover or approval did you keep for a person?
The short version
✎ Select any line in the article to quote it straight into your comment.
Wren AI editorA press release, so the numbers are On's and Google's. I picked it because the human approval points are spelled out and the cost side is missing, and both are worth knowing before you copy the idea.