Back to blog Perspectives

Your Cloud, Your Keys, Your Agents: What Stays Yours on an Agent Platform

Phinite teamNov 5, 20263 min read

Every platform promises openness. The useful question is more specific: when you leave, or when a regulator asks, what is still yours? For an agent platform the answer hides in four places. Here is what each one should mean, and how to check.

Where it runs

Bring your own cloud should mean the platform deploys inside your AWS, Azure or Google Cloud account, Kubernetes-native, with your data, prompts and traces never leaving that boundary. Sovereign deployment extends the same idea to your own data centre, air-gapped, with no external dependencies.

The check: ask where a trace is stored and who can read it. If the answer involves the vendor's account, it is not your cloud.

Whose model keys

Bring your own keys should mean your contracts with model providers, your rate limits and your spend. Closed models through your keys. Open-weight models hosted by you, on GPUs you control, when the data or the economics call for it. Managed inference where it is good enough, private inference where it is not.

Routing is the part people forget. A platform that routes each task to the best model is only open if you can change the model later without rebuilding the agents. Versions should record which model made each decision, so a provider's silent upgrade is visible in the trace rather than discovered in production.

Which agents it accepts

A surprising amount of lock-in comes from the builder. If the only agents the platform governs are the ones built in its own editor, every existing LangGraph or CrewAI agent is either rebuilt or ungoverned. Bring your own agents should mean those agents run on the same runtime, behind the same gate, with the same trace and evaluations, no rewrite required.

The check: bring one existing agent to the demo and ask to see it pass through the policy check.

Whose guardrails it enforces

Many organizations already own content-safety profiles in AWS Bedrock Guardrails, Azure Content Safety or Google Model Armor, mandated by a cloud or security team. Bring your own guardrails should mean attaching that profile to an agent, testing the connection, and letting it run on every turn, rather than rebuilding the same rules in a second system with a second owner.

Why openness and governance go together

It is tempting to see "open" and "governed" as a trade-off: the more the platform controls, the more it locks you in. We think the opposite is true. A platform that governs every action through one gate, records every decision with its version, and runs in your boundary is the one that is easiest to leave, because everything that matters, the policies, the evaluations, the traces and the agents themselves, is in a form you own and can read.

The operating layer should be the thing you keep. The vendor should be the thing you can change.

A short checklist

  • Deployment: your cloud account or your data centre, with traces inside the boundary.
  • Keys: your model contracts, with the model version recorded on every decision.
  • Agents: existing LangGraph and CrewAI agents run without a rewrite, behind the same gate.
  • Guardrails: your cloud provider's profile attaches and runs on every turn.
  • Exit: policies, evals and traces exportable in a form your next platform can read.

Ask for all five in the first conversation. The answers tell you more about the platform than any feature list.

Written byPhinite teamWrites on building, running and governing AI agents in production.

Prove what your agents changed in the business.