Your existing cloud decides this, not us
Technology
Cloud choice is rarely a technical argument. The team that already has the IAM policies, the billing account, and the security review behind them ships months before the team starting that conversation from zero.
So we build on both. On Google Cloud we run a nine agent Gemini pipeline serving 800 to 1,000 developers a day, managed as infrastructure as code across separate dev, staging, and prod environments. We build the same class of system on AWS.
Running
What it takes to run agents on Google Cloud
- The whole system lives in one account
Vertex AI Agent Engine, Cloud Run, Cloud Build, Artifact Registry, Secret Manager, BigQuery, and Cloud Logging carry a nine agent pipeline serving 800 to 1,000 developers a day. Terraform manages dev, staging, and prod as separate environments, with automated CI/CD gates between them.
- Programmatic agents where a model is waste
Not every stage of an agent pipeline should call a model. Two of ours do not. Deterministic validation and a lookup against a canonical registry are cheaper, faster, and cannot hallucinate, which is a better property than a well written prompt. Reach for code where code is sufficient.
- Quota becomes the failure mode at scale
As traffic grew, API quota errors became the primary way the system failed. The answer was a three layer retry strategy and graceful fallback to the last valid output, not a bigger quota request. The experience shifts from request failed to request takes longer.
- The MCP server gets its own project
Three of the nine agents query an indexed knowledge base at runtime over MCP. Retrieval sits in a separate project so it draws on a separate quota from the model calls, and results are capped per query, which stops the context window from growing without bound.
- Five phases, all of them after launch
The pipeline running today is the fifth version. Each one answered something we watched happen in production: token cost, evaluation latency, quota errors, redundant work. Budget for the versions that come after the launch. On a system this size, the first deploy is a starting position.
- When Google Cloud is the wrong call
If your IAM, billing, and security review already live in AWS, opening a GCP project means governing a second cloud. We build the same class of agent on AWS with Bedrock, AgentCore, and Connect. The platform that is already approved usually wins.
