Signal

Why it matters: AWS put its original agent platform into indefinite maintenance mode with no end-of-life date and no migration deadline — and the missing deadline will cost people more than the deprecation would have.

Amazon Bedrock Agents — which AWS's own documentation dates to a November 2023 launch — is now Amazon Bedrock Agents Classic, and has been closed to new customers since July 30, 2026. Read strictly as a deprecation, the maintenance-mode documentation is close to the most generous version available: existing agents keep working, every API stays available, existing CloudFormation and Terraform templates continue to work for allowlisted accounts, there is no charge for the service itself, and AWS states plainly that there is no planned end-of-life date and no migration deadline. Thirty-two months from launch to maintenance mode is quick, but nothing was taken away from anyone already running on it.

One line further down the same page matters more than all of that. The model catalog available in Classic is frozen as of July 30, 2026, and models released after that date are reachable only through AgentCore. No new features are planned either.

That is the point where an agent orchestration layer stops behaving like ordinary infrastructure. A queue or a database in maintenance mode is fine, and arguably improved by it — frozen means predictable. An orchestration layer exists precisely because it sits between your code and a model that gets better on a cadence measured in months. Freeze its catalog and the component whose job was to deliver that improvement becomes the one component guaranteeing you cannot have it. Nothing fails. The agent simply falls further behind at whatever rate the frontier moves, and the cost arrives as mediocre output with no incident to attribute it to.

The strongest case for AWS's approach deserves stating properly, because it is a good one. The re-architecture is defensible on the merits — separating runtime, memory, gateway and identity is a better factoring than a single managed agent object, and a great deal was learned between 2023 and 2026. AWS also shipped real migration tooling rather than a blog post: an amazon-bedrock skill in the agent toolkit for AWS that drives a coding assistant through inspecting the existing agent, assessing eligibility and scaffolding the replacement, and which the documentation says never modifies the source agent. An open-ended support commitment costs a vendor something real. Most deprecations are worse than this one.

But the generosity is exactly what removes the forcing function. A hard sunset date is unpleasant and it gets budgeted, because a calendar entry is something a quarter can be planned around. "Nothing breaks, no deadline" never becomes any quarter's priority, while the gap compounds quietly. The migration is not free either, and the same documentation is candid about that: stage-specific prompt overrides are "not directly replicated," and routing-mode multi-agent is "not straightforward today." Continued access is conditional rather than owned — allowlisting is per-account and determined automatically by whether that account used Bedrock Agents in the previous twelve months, there is no exception process, and accounts outside the list get a 403 from CreateAgent.

The lesson is not to avoid vendor agent platforms, which have no shortage of compensating advantages. It is that there is a specification question worth asking during procurement that almost nobody asks: what is this layer's model-adoption latency — how long between a model shipping and this layer being able to call it? Through 2024 and 2025 that number stayed invisible because it was small. For Classic it is now unbounded, and it was disclosed in an FAQ answer rather than on a pricing page. It is also why the machinery behind this site stays aggressively boring: a Signal edition is a row in a database rendered by a static build, and the pricing comparator prints a verification date beside every figure. There is no orchestration layer between the content and the page, so there is nothing here that can be frozen — a much smaller problem than the one AgentCore is solving, not a better answer to it.