Part 27: Where should the agent runtime live?

Part 27 · One Agent, Five Platforms
Part 26 put the five builds side by side. This part asks the question behind them: where should the agent runtime live? In your integration platform, with a hyperscaler, with your ERP vendor, or on a platform you run yourself?
Like Part 20, this is a discussion part, so there are no new platform facts. It draws on what the five documentation-only builds in Parts 11 to 25 showed. I'm running a poll on LinkedIn alongside it, and I'll share the results in a later part.
The four options
1. Your iPaaS (MuleSoft, Boomi)
The agent runs next to your integrations. In Part 11, MuleSoft put the tools in Mule flows and the guardrails in the same gateway as the APIs. In Part 15, Boomi did the same with AI Gateway.
- Best when: most of what the agent touches already goes through the iPaaS, across several ERPs and SaaS systems.
- Watch out for: the builds found less on distinct agent identity and built-in evaluation here than on the hyperscalers. Regions and runtime options vary too: Boomi listed no EU runtime cloud.
2. A hyperscaler (Azure, AWS)
The agent runs in your cloud account. Part 22 and Part 25 had the most documentation behind them: a directory identity per agent, OpenTelemetry tracing, built-in evaluation and versioned deployments.
- Best when: your AI and data platforms are already there, and you want the agent tooling to keep moving fast.
- Watch out for: the approval and its audit trail are assembled from parts on both. Foundry's MCP tool calls on Azure time out after 100 seconds, and AWS's own approval tutorial skips authentication.
3. Your ERP vendor (SAP)
The agent runs where the data and the approvals already are. Part 18 found SAP's flexible workflow was the only approval that lives in the system that posts the invoice.
- Best when: one ERP dominates, and your approvers already work in its inbox.
- Watch out for: the vendor's rules of access shape the design (Part 17), and several capabilities were still on the roadmap. The fewer systems outside the ERP, the better this fits.
4. Your own platform
An open-source agent framework on infrastructure you run. This series didn't build one, so this is the one option without evidence from the builds.
- Best when: you need full control over models, data location and release cadence, and you have a platform team to run it.
- Watch out for: everything the other four provide, you build: identity, guardrails, tracing, evaluation and upgrades.
What the builds actually showed
Two things held on every platform, and they make the choice less final than it looks:
- The contract travelled. The nine tools from Part 7 mapped onto all five without changing the contract.
- The gate stayed outside the runtime. The
approval_idcheck lived in the process API every time, as Part 19 argued. Wherever the agent ran, it never became the last check.
So the runtime is replaceable, as long as the tools sit over process APIs and the controls sit outside the agent. What's harder to move is the approval experience and the identity model, because people and auditors get used to them.
Questions that decide it
Rather than a ranking, four questions:
- Where do your approvers already work? An ERP inbox, Teams, or a separate tool.
- Where must the data stay? Region, network and whether model calls can leave your account.
- Who will run it? An integration team, a cloud team, the ERP team, or a platform team.
- How many systems does the agent touch? One ERP, or many systems across vendors.
Where I stand
For this series' fictional company, I'd keep the tools and the gate in the integration layer, and choose the runtime by where approvers work and who will operate it. But that's a reasoned guess from documentation, not a result.
The poll
On LinkedIn, I'm asking one question with four answers:
- iPaaS: next to your integrations.
- Hyperscaler: in your cloud account.
- ERP vendor: next to the system of record.
- Own platform: a framework you run yourself.
Where would you run your agents?
This series describes a reference build on a fictional company. This part is a discussion based on the documentation-only builds in Parts 11 to 25; no hands-on build has been run yet. Views are my own.
原文:https://www.a8q9.cn/en/article/part-27-where-should-the-agent-runtime-live-ebd02a.html