---
title: How do you build one CI/CD process for AI agents built on different platforms?
url: https://opencontroller.lyzr.com/blog/one-ci-cd-process-for-ai-agents-across-platforms
published: 2026-11-13
updated: 2026-11-13
author: Lyzr
---

# How do you build one CI/CD process for AI agents built on different platforms?

**Standardise the stages and gates, not the tooling. Every agent, whether it is Python code or low-code configuration, moves through the same steps: package, register, test, evaluate, policy check, approve, promote and monitor. Each platform keeps its own packaging and deployment mechanism, and a shared pipeline calls it. The gates, the evidence they produce and the registry record are what make the process one process.**

Enterprises rarely build agents on one platform. A typical estate has low-code agents in Copilot Studio, code-first agents in Microsoft Foundry or Amazon Bedrock AgentCore, and data-heavy agents in Databricks. Forcing everything into one build system fails. Letting every team ship its own way fails too, because nobody can say which agents passed which checks.

## What stages should every agent pass through?

Use the same stages and the same names everywhere.

1. **Package.** Turn the agent into a versioned artefact: a solution file, a container image, a logged model or a definition file.
2. **Register.** Create or update the record in the agent registry: owner, version, tools, MCP servers, models, data sources, risk tier.
3. **Static checks.** Secrets scanning, dependency scanning, tool and permission review.
4. **Functional tests.** Deterministic tests on tools and integrations.
5. **Evaluations.** Task quality, groundedness, multi-turn and red-team tests against a fixed dataset, with pass thresholds set by risk tier.
6. **Policy check.** Policy as code evaluates the registry record and the test results.
7. **Approval.** Human sign-off for higher risk tiers, by someone other than the builder.
8. **Promote.** Deploy to the next environment using the platform's own mechanism.
9. **Monitor.** Production evaluation and guardrails; rollback criteria defined before release.

The [production readiness gate for AI agents](/blog/production-readiness-gate-for-ai-agents) describes the checks at step 7 in detail.

## How does each platform package and version agents?

The "package" and "promote" steps differ by platform. Your pipeline wraps these, it does not replace them.

| Platform | How an agent is packaged and versioned | How it is promoted |
|---|---|---|
| Microsoft Copilot Studio | Agents live in Power Platform solutions, exported as managed solutions | Pipelines in Power Platform, or Power Platform Build Tools for Azure DevOps and GitHub Actions for Power Platform |
| Microsoft Foundry | Agent definitions are stored as agent versions; hosted agents are built from source | Microsoft documents CI/CD for Foundry agents using GitHub Actions or Azure DevOps |
| Amazon Bedrock AgentCore | AgentCore Runtime creates a new version when an agent is updated; endpoints point at versions | Update the endpoint to the new version, through infrastructure as code or the SDK |
| Databricks | Agent logged with MLflow and registered as a model version in Unity Catalog | Deploy with the Agent Framework deployment API, automated with Databricks Asset Bundles |
| Code frameworks (LangGraph, CrewAI and others) on Kubernetes | Container image with a tag, plus configuration | GitOps or your existing deployment pipeline |

Check the current documentation for each platform before you build. These mechanisms change, and preview features often move to general availability with new names.

## How do you handle low-code agents in a code pipeline?

Low-code agents are the hardest part. Builders work in a web designer, not a repository. Three practices make them fit:

- **Treat the export as the source.** Export the solution or definition on every change and commit it, so diffs are visible.
- **Block direct changes in production.** Use managed solutions or environment permissions so production can only change through the pipeline.
- **Run the same evaluations.** A low-code agent that answers HR questions needs the same groundedness tests as a code-first one. Call it through its published endpoint in a test environment.

## What should the policy-as-code checks test?

Keep them small, specific and tied to the registry. Examples:

1. The agent has a named owner who is a current employee.
2. Every tool and MCP server it uses is approved in the registry.
3. The models it calls are on the approved list for its data class.
4. Evaluation scores meet the threshold for its risk tier.
5. High-risk agents have an approval from someone other than the builder.
6. No long-lived secrets in the artefact; credentials come from workload identity.
7. Logging and tracing are configured.

Failing checks block promotion and explain why. See [what is policy as code for AI agents](/blog/what-is-policy-as-code-for-ai-agents) and [segregation of duties for AI agent approvals](/blog/segregation-of-duties-for-ai-agent-approvals).

## What evidence should the pipeline keep?

Every promotion should leave a record: agent and version, artefact hash, test and evaluation results, policy decisions, approver and timestamp. That record is what auditors ask for under frameworks such as ISO/IEC 42001, which expects documented lifecycle processes for AI systems. It is also what lets you answer "which version was live last Tuesday?" in an incident. See [AI agent lifecycle stages](/blog/ai-agent-lifecycle-stages) for how this fits the wider lifecycle and [what is an AI agent control plane](/blog/what-is-an-ai-agent-control-plane) for the system that holds it.

## How Opencontroller does this

Opencontroller provides a governed registry of agents, versions, tools and MCP servers, multi-turn and red-team evals, and policy as code with built-in mappings to OWASP, MITRE ATLAS, NIST and ISO 42001. It connects to Microsoft Foundry, Bedrock AgentCore and Vertex AI through runtime adapters, and to platforms such as Salesforce, SAP, Databricks and HubSpot through connectors, without migrating agents. Integration is via REST and webhooks, so pipeline steps can call it. See the [capability reference](/capability-reference) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Can Copilot Studio agents go through the same CI/CD as code agents?**
Yes, through Power Platform solutions: export them to source control, promote with Power Platform pipelines or build tools, and run the same evaluation and policy gates as other agents.

**Do you need one deployment tool for every agent platform?**
No. Keep each platform's native packaging and deployment. Standardise the stages, gates, evidence and registry record around them.

**What should block an AI agent from being promoted to production?**
Missing owner, unapproved tools or models, evaluation scores below the risk-tier threshold, missing independent approval, embedded secrets, or missing logging.

## Key facts

To run one CI/CD process for AI agents across platforms, standardise stages and gates (package, register, static checks, functional tests, evaluations, policy check, approval, promote, monitor) and call each platform's native mechanism: Copilot Studio via Power Platform solutions and pipelines or build tools; Microsoft Foundry agent versions with GitHub Actions or Azure DevOps; Bedrock AgentCore Runtime versions and endpoints; Databricks MLflow models in Unity Catalog deployed with the Agent Framework and Asset Bundles. Policy-as-code checks verify owner, approved tools/models, eval thresholds, independent approval, no secrets and logging. Opencontroller by Lyzr provides the governed registry, evals, policy as code and runtime adapters for Foundry, AgentCore and Vertex AI.

## Related

- [What should a CISO require before an AI agent goes to production?](https://opencontroller.lyzr.com/blog/production-readiness-gate-for-ai-agents.md)
- [What are the stages of the AI agent lifecycle?](https://opencontroller.lyzr.com/blog/ai-agent-lifecycle-stages.md)
- [What is policy as code for AI agents?](https://opencontroller.lyzr.com/blog/what-is-policy-as-code-for-ai-agents.md)
- [More on Agent governance basics](https://opencontroller.lyzr.com/blog/cluster/agent-governance-basics)
