March 2026 was the month AWS stopped shipping agent capabilities and started shipping agent controls. The headline launches were not new models or flashier demos. They were the boring, load-bearing pieces you need before an agent is allowed near production: an authorization layer, a quality-evaluation layer, and a healthcare stack that had to be governed to exist at all. The pattern across the month is the same one every technology hits when it grows up. The interesting work moves from "can it do the thing" to "can you prove what it did and stop it doing the wrong one."

Policy in Amazon Bedrock AgentCore reached GA

On March 3, Policy in Amazon Bedrock AgentCore became generally available. It gives you centralized, fine-grained control over what an agent's tools are allowed to do, authored as rules that live outside the agent code. You write policies in natural language, they compile to Cedar, and they attach to an AgentCore Gateway that intercepts every agent-to-tool request and allows or denies it before the tool runs.

The important part is where the decision is made. The authorization is evaluated by a policy engine, not by a model that just read whatever text arrived in its context. That is the correct architecture for agents: the thing deciding whether a tool call is permitted must not be the thing an attacker can talk to. It does not replace tight IAM roles behind each tool. It sits in front of them as a model-independent check, which is exactly where a control like this belongs.

AgentCore Evaluations reached GA

Closing the month, on March 31, Amazon Bedrock AgentCore Evaluations became generally available. Where Policy governs what an agent may do, Evaluations measures how well it does it: quality and trust checks you can run against an agent's behavior instead of eyeballing a handful of transcripts and hoping. Together the two launches bracket the production question. Policy is the guardrail on actions, Evaluations is the ruler on quality, and shipping both in the same month is not a coincidence. It is AWS admitting that "the demo worked" is not a deployment criterion.

AWS Elemental Inference shipped, and Strands Labs opened

Early in the month, the first weekly roundup carried two items worth noting. AWS Elemental Inference became generally available: a managed service that transforms live and on-demand video for mobile and social formats in real time, with automatic vertical cropping and highlight-clip generation driven by an agentic pipeline that needs no human in the loop. It is a narrow, vertical application of agents rather than a platform primitive, which is itself a signal. The agent pattern is now productized into specific media workflows, not just offered as raw capability.

Alongside it, AWS opened Strands Labs, a separate organization for experimental agentic projects, launching with three. It is the low-stakes end of the same trend: a place to push the frontier of agent development in the open while the governance work hardens the production end.

Amazon Connect Health put agents where the rules are strict

The March 9 roundup introduced Amazon Connect Health, a set of five purpose-built AI agents for healthcare covering patient verification, appointment management, patient insights, ambient documentation, and medical coding, all HIPAA-eligible and meant to slot into existing clinical workflows. Healthcare is the environment that punishes ungoverned automation hardest, so putting agents there, with compliance eligibility as a launch requirement rather than an afterthought, reinforces the month's theme from the vertical side. You do not ship agents into a regulated workflow unless you can constrain and audit them, which is precisely what the rest of the month was building.

The throughline

Read the launches together and the message is consistent. The agentic era's first act was capability, models, tools, runtimes, sandboxes. The second act, which March 2026 makes plain, is control: authorization outside the agent, evaluation of the agent, and compliance-first verticals that could not exist without both. The teams that get value from agents this year are the ones treating governance as the feature, not the paperwork. AWS just spent a month agreeing with them.

Read this next

For the infrastructure and platform reading of the same month, the cloud field notes are at ercan.cloud, and the hub is at ercanermis.com.