Dev.to AI 🤖 Ai 👁 0 📖 5 min read

Part 30: Identity and governance for agents across five vendors

Part 30 · One Agent, Five Platforms Five platforms means five policy models, unless you design one. Each vendor has its own idea of who an agent is, where its permissions live, what gets logged and where the agent list

Part 30: Identity and governance for agents across five vendors

Part 30 · One Agent, Five Platforms

Five platforms means five policy models, unless you design one. Each vendor has its own idea of who an agent is, where its permissions live, what gets logged and where the agent list is kept. If you run agents on more than one, governance has to sit above them.

This part puts the five builds side by side on six controls, and then maps one policy model onto all five. It draws on the documentation-only builds in Parts 11 to 25. On 18 November 2026 I re-checked the two facts that had moved most: the AWS and Microsoft agent registries.

Governance matrix: six controls across five platforms

1. Agent identity

Part 23 set out the two ways an agent can act: as itself, with its own identity and permissions, or on behalf of a user, carrying a person's identity. The builds found very different levels of support for the first:

  • Azure: each new Foundry agent gets its own Entra agent identity, which starts with no permissions (Part 22).
  • AWS: a workload identity in AgentCore Identity, plus an IAM execution role with a permission boundary (Part 25).
  • SAP: SAP Cloud Identity Services, with token exchange so the agent gets a scoped credential rather than passing the caller's token on (Part 18).
  • MuleSoft: gateway authentication and attribute-based access by claims. A distinct identity per agent wasn't clear from the pages I read (Part 11).
  • Boomi: agent keys per subscription. I found no distinct agent identity (Part 15).

The policy: every agent gets its own identity, owned by a named team, even if the platform makes you build it from a service principal or workload identity.

2. Least privilege per tool

An identity alone isn't enough. The invoice agent can read invoices but can't post adjustments without an approval_id. Where that limit is enforced differs:

  • MuleSoft: tool allow and block rules in Omni Gateway, covered in Series 1 Part 5.
  • AWS: Cedar policies at AgentCore Gateway.
  • Azure: tokens issued for one audience at a time, with RBAC at the narrowest scope.
  • Boomi: each agent imports its own tool list from the MCP Registry.
  • SAP: I found no way to give one agent a narrower tool set than another.

The policy: list each agent's tools and scopes in one place. Enforce it at the gateway where the platform allows it, and in the process API either way. The approval_id check from Part 6 stays in the process API on every platform, so no vendor setting can remove it.

3. Audit trails that survive platform boundaries

In Part 28, one question went from Azure to MuleSoft. That's two platforms, two log stores and two trace formats for one case. An auditor will want to follow it end to end.

What the builds found:

  • MuleSoft: OpenTelemetry traces across agents, brokers and MCP servers, with export.
  • Azure: OpenTelemetry traces in Application Insights, plus Entra sign-in and audit logs you can filter by agent.
  • AWS: OpenTelemetry telemetry in CloudWatch. CloudTrail audit coverage applies to Gateway and the Agent Registry only.
  • Boomi: session traces with tool calls, tokens and latency. No export documented.
  • SAP: gateway logs, traces and analytics. The data, the approval and the posting are in one system.

The policy: one case id that travels in every A2A message and tool call, OpenTelemetry wherever it's offered, and a central store for the decisions that matter: who approved what, when, and which agent acted. Where a platform can't export, record those decisions in the process API instead.

4. PII and data residency

PII. MuleSoft has an LLM PII Detection policy in Omni Gateway. SAP lists data masking and content filtering. Boomi's agent guardrails have word filters and custom regex, but no separate PII feature. I didn't check PII features on Azure or AWS during the builds, so that's open.

Residency. The builds found:

  • MuleSoft: agent networks on the US and EU cloud hosts only, or a customer-managed runtime with Runtime Fabric.
  • Boomi: four managed runtime clouds, in the US, UK, Japan and Australia. No EU cloud listed. The Flex runtime is in early access.
  • Azure: virtual network isolation, bring-your-own storage and a self-hosted API Management gateway.
  • AWS: VPC networking for AgentCore Runtime.
  • SAP: nothing in the sources I could read.

On Azure and AWS, I didn't check where model processing happens for each model.

The policy: classify the data each tool returns, and decide where each class may go before choosing where the agent runs. An agent that reads supplier bank details has fewer places it can live.

5. A central registry of agents and tools

Every vendor now has some kind of registry. None of the pages I read promises one list for all five:

  • AWS: AWS Agent Registry is a catalog for "MCP servers, tools, agents, agent skills, and custom resources", with approval settings so a curator approves records before they can be found. It became generally available on 31 August 2026. The pages don't say whether it covers agents on other clouds.
  • Microsoft: registry experiences are converging under Agent 365, described as "a unified inventory" covering "both Microsoft and non-Microsoft agents". Entra keeps the identities.
  • MuleSoft: the Agent Fabric registry.
  • Boomi: the MCP Registry for tools, and Agent Control Tower, which manages third-party agents.
  • SAP: I didn't find an agent registry in the pages I read.

The policy: pick one registry of record. Every agent and every tool is listed there with an owner, its identity, its tools and scopes, its data classes, its region and its agent card from Part 29. The vendor registries follow it, not the other way round.

6. One policy model, mapped to each vendor

Put together, the policy model is short, and it's the same on every platform:

  • Identity: one identity per agent, with an owner. Enforced by Entra agent ID, AgentCore workload identity, SAP Cloud Identity, or a service identity on MuleSoft and Boomi.
  • Tools: a list of tools and scopes per agent. Enforced by gateway rules where available, and by the process API always.
  • Approval: an approval_id for every write. Enforced in the process API, the same way on all five.
  • Audit: a case id, the decisions and who approved them. Kept through OpenTelemetry export, or in the process API where there's none.
  • Data: data classes and the regions each may go to. Enforced by the runtime region, network isolation and PII filters.
  • Registry: one registry of record. Vendor registries are synced from it.

Part 4 said the protocols underneath are shared: OAuth 2 and OpenID Connect for identity, MCP for tools and A2A between agents. The governance can be shared too, as long as it's defined above the vendors and only enforced in them.

What I'd take from this

The vendors differ most on identity and audit, and least on what you'd want from them. Write the policy once, keep the approval and the audit record in your own process APIs, and treat each platform's controls as one way of enforcing it.

Who owns agent governance in your organisation?

This series describes a reference build on a fictional company. This part is based on the documentation-only builds in Parts 11 to 25 and on AWS and Microsoft registry pages read on 18 November 2026; no hands-on build has been run yet. Views are my own.

📰 Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.