Bedrock Agents maintenance mode: your agent platform froze

If you built anything on Amazon Bedrock Agents, the product you bought no longer exists under that name. AWS renamed it Amazon Bedrock Agents Classic and put it into maintenance mode as of July 30, 2026. Nothing broke, nothing shut off, and there's no end-of-life date. That's exactly why it's easy to miss the part that matters.

What actually happened

Bedrock Agents launched in November 2023 as the managed way to run an agent on AWS: you declared a model, action groups, and a knowledge base in the console, and AWS owned the orchestration loop. As of July 30, it's no longer open to new customers. The allowlist is automatic and per-account — if an AWS account had Bedrock Agents activity in the previous 12 months, it keeps full access. Accounts without prior usage get an AccessDeniedException (HTTP 403) on CreateAgent and InvokeInlineAgent. AWS states there is no exception process.

Existing agents keep running. Every other API — InvokeAgent, UpdateAgent, action groups, aliases — stays available. Knowledge Bases, Guardrails, and Bedrock model inference are untouched. Pricing doesn't change; Agents Classic never carried its own charge.

Two lines in that FAQ are the actual news. First: the model catalog is frozen as of July 30, 2026. Models released after that date go to AgentCore, not to Agents Classic. Second: no new features are planned. AWS points migrations at Amazon Bedrock AgentCore, either through a config-based managed harness or by deploying code-defined agents on AgentCore runtime with Strands, LangChain, the OpenAI Agents SDK, the Claude Agent SDK, or your own framework.

Why it matters for your business

"No end-of-life date" reads like you have time. The frozen model catalog says otherwise. A managed agent whose model list stops advancing gets worse every quarter relative to everything around it — not because AWS broke it, but because the rest of the market moved and your orchestration layer can't follow. That's a deadline with no date on it, which is harder to plan around than a hard cutoff.

The second-order problem is reproducibility. Your Terraform and CloudFormation still work in allowlisted accounts. Stand up a new AWS account for a new client, a new environment, or a spun-out entity, and the same templates fail with a 403. Infrastructure that can't be recreated from code in a fresh account isn't infrastructure, it's an artifact.

The structural lesson is the one we keep landing on: the fragile layer in an agent stack is the managed orchestration loop, not the model and not the tools. Models are swappable, tools are yours, and the loop is the piece the vendor owns and can freeze. AWS's own migration table is honest about the gap — stage-specific prompt overrides don't carry across, multi-agent routing isn't straightforward, and the built-in AMAZON.UserInput elicitation becomes an explicit tool you define yourself. Every one of those is a place where you wrote configuration instead of code, and now you write the code anyway.

Sources: AWS — Bedrock Agents Classic maintenance mode, AWS — Automate tasks using AI agents.