Building an AI agent that works in a demo has become easy. Running one safely in production, inside a real organisation, is still the hard part. In 2026, both major clouds have moved to close that gap with managed agent runtimes.
What has changed
On AWS, the Amazon Bedrock AgentCore harness became generally available on 18 June 2026. Teams define an agent with CreateHarness and run it with InvokeHarness, without writing orchestration code or building a container by default. Each agent runs in an isolated sandbox, with managed memory, CloudWatch traces, evaluation, versioned endpoints and tool connectivity through AgentCore Gateway or remote MCP servers.
On Azure, Microsoft Foundry Agent Service reached general availability for its core service in March 2026, positioned as a governed, production-ready platform with security isolation, life-cycle control and cost transparency. Microsoft has also been retiring older building blocks: the Assistants API was scheduled to retire on 26 August 2026, with the Responses API and Foundry Agent Service as the replacement.
What the platforms do not decide for you
A managed runtime removes a lot of plumbing. It does not decide which workflow is worth automating, what an agent should be allowed to touch, or how you will know it is getting answers right. Those decisions are where most agent projects succeed or stall.
Six checks before an agent goes live
- Start with one well-defined workflow. Pick a task with a clear owner, clear inputs and a measurable result.
- Give the agent least-privilege access. Each tool and API should carry only the permissions that the task needs, using the same identity controls as the rest of your estate.
- Build an evaluation set first. Test suites and regression checks should run before every release, not after something goes wrong.
- Trace every run and track cost per task. You need to see what the agent did, why, and what it cost.
- Keep a human in the loop for high-impact actions. Approvals for payments, deletions and customer-facing messages are cheap insurance.
- Make ownership explicit. Someone on your team should hold the runbook, the alerts and the budget.
How we approach it
We usually start with a four to six week proof of concept on one workflow, then harden it for production with a security review, an evaluation pipeline, monitoring and a runbook your team can own. You can read more on our Agentic AI & AI Agents page, or get in touch to talk through your first use case.
Sources: AWS Amazon Bedrock AgentCore harness general availability (18 June 2026); Directions on Microsoft Foundry Agent Service roadmap; Microsoft Learn Foundry documentation.
