Back to blog Perspectives

Build or Buy an Agent Platform? Seven Questions Your Risk Committee Will Ask First

Phinite teamNov 5, 20263 min read

The build-versus-buy conversation about agent platforms usually happens between engineers, and it usually ends with a framework. LangGraph or CrewAI, a vector store, a gateway, some dashboards. The conversation that decides whether any of it reaches production happens later, in a different room, with risk, audit and compliance.

We think the second conversation should come first. Here are the seven questions it will contain. Whether you build or buy, the platform has to answer every one, and the answers have to be reproducible by someone other than the engineer who set it up.

1. Who approved it?

Every agent in production needs a named owner and a record of who approved the version that is live. Not a Slack thread. A record, attached to the build, that survives the engineer leaving.

2. What data can it access?

Scoped credentials per agent, per environment, with the scope visible to someone who did not write the code. "It uses the service account" is the answer that fails this question.

3. Which model and version decided?

Each decision has to be traceable to the exact model, prompt, tools and knowledge sources that produced it. Model providers change versions underneath you; the trace has to capture what actually ran, not what was configured last quarter.

4. Which tools can it invoke?

This is the policy question. Allow, hold for a person, or deny, for every tool, integration, MCP server and code-runner action, with argument constraints where they matter: amount limits, allowlists, required fields. Default deny, with the allowlist simulated before go-live.

5. Can every action be traced?

Not just the model calls. Every tool call, data read and message, tied to a case, a version and the policy that permitted it. If an auditor picks a case at random, the full path from intake to action should open in minutes.

6. Are policies enforced at runtime?

A policy that lives in a document is not enforcement. The check has to happen before the action runs, on every action, whoever built the agent and whichever framework it came from. If a team can route around the gate by calling the model directly, the answer to this question is no.

7. Can we reproduce and roll it back?

Versioned releases, separate environments, and a single move to roll back to the last good build. Plus a way to replay a case against a version to show what would have happened.

What this means for build versus buy

Notice that none of the seven questions is about the quality of the agent. They are about operating it. Teams that build often answer the first three well and discover the last four under pressure, usually after an incident, usually one tool at a time: a gateway here, a logging pipeline there, an approval workflow in a ticketing system that nobody connects to the trace.

Buying does not automatically answer them either. A platform that sells seats and dashboards can fail question six just as badly. The test is whether policy, tracing, evaluation and versions are one system with one permissions model, or five vendors stitched together.

There is also a middle path that most teams miss. Bring your own agents. If your engineers have already built in LangGraph or CrewAI, those agents can run on an operating layer that answers the seven questions, with no rewrite, behind the same gate as everything else. The build-versus-buy choice stops being about the agent and becomes about the layer underneath it, which is where it belonged all along.

Run the test before you decide

Whichever way you lean, ask for one live demo on one real process and work through the seven questions in order. Open a case. Show the versions. Show the policy that allowed the action. Try to roll back. The platform that can do all seven in front of you is the one your risk committee will sign off on. The other kind is the one that gets stuck in pilot.

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

Prove what your agents changed in the business.