Through June and July 2026 AWS moved a broad set of services into maintenance mode and pointed customers toward newer replacements. Several of them sit right in the AI stack: enterprise search, the original Bedrock agents feature, and the Amazon Q assistants. None of the AI services is being shut down overnight, but each is now a service you should plan to migrate off rather than build new work on. One correction up front, because the wording matters: Amazon Q Business is not “dead.” It is closed to new customers and still fully supported.

A split image: dark tangled cables and a server rack on one side, clean green modular cubes on the other, representing migration from a legacy system to a modern one.
A lifecycle wave is a migration prompt, not a fire. Maintenance mode means no new features and no new customers, while existing workloads keep running.

What happened

AWS placed the following services into maintenance mode or end of support. “Maintenance mode” is AWS’s term for a service that keeps running with security and bug fixes but gets no new features and, past a cutoff date, no new customers.

ServiceWhat changedKey dateAWS points you to
Amazon KendraMaintenance mode30 June 2026Bedrock Managed Knowledge Base
Amazon Bedrock AgentsRenamed “Classic”, closed to new customers30 July 2026Bedrock AgentCore
Amazon Q BusinessClosed to new customers, still supported30 July 2026Amazon Quick
Amazon Q Developer (IDE plugins, paid tiers)End of support30 April 2027Kiro
Amazon Cognito SyncClosed to new customers30 July 2026AppSync, DynamoDB
Amazon SageMaker Studio LabClosed to new customers30 July 2026SageMaker Studio
AWS ProtonFull end of support, data deleted after7 October 2026CloudFormation Git Sync
Amazon Mechanical TurkFull permanent closure — existing customers affected, not just new30 September 2026No direct successor named

Two of these need precise wording. Amazon Q Business is, in AWS’s own language, “no longer open to new customers” as of the same wave, with AWS’s June 2026 announcement giving July 30, 2026 as the effective date (AWS’s Q Business migration guide itself states the policy but does not repeat the date) — and Q Business “remains fully supported,” with no shutdown date announced; AWS recommends new work go to Amazon Quick. Amazon Q Developer is a separate product with a separate timeline: its IDE plugins and paid subscriptions reach end of support on 30 April 2027, with Kiro as the successor, while Amazon Q in the AWS Console, the documentation, and the Slack and Teams integrations are unaffected.

The clearest picture is one consolidation into three newer platforms plus one editor: retrieval and search move to Bedrock Managed Knowledge Bases, agent building moves to Bedrock AgentCore, assistants and business intelligence move to Amazon Quick (AWS’s marketing name for the service; it is not called “Quick Suite” in AWS’s own docs), and the AI coding assistant moves to Kiro. Amazon Bedrock itself, its Knowledge Bases, and Guardrails are not affected; only the original 2023 “Bedrock Agents” feature, now labelled “Classic,” is.

Only AWS Proton is a true shutdown: support ends 7 October 2026, after which the console and resources become inaccessible and the data is deleted, though CloudFormation stacks Proton deployed stay intact.

The SageMaker AI cluster

The same wave swept up a cluster of Amazon SageMaker AI features, the machine-learning tooling many teams use to train, monitor, and label. All of the following were announced on 30 June 2026 and close to new customers on 30 July 2026, in the same maintenance-mode sense as above: Clarify (bias and explainability), Model Monitor (drift detection), Debugger, Ground Truth (data labeling), Augmented AI (A2I) (human review), Role Manager, and Geospatial capabilities. The Mechanical Turk workforce integration used by Ground Truth and A2I is a separate, more severe case: AWS announced on 25 August 2026 that Mechanical Turk itself is being permanently closed on 30 September 2026 — not just closed to new customers, but shut down for everyone, which removes the Mechanical Turk workforce option from both Ground Truth and A2I entirely.

For the two most widely used, AWS names specific replacements. For Clarify, it points to Amazon Bedrock Evaluations and Guardrails plus the open-source fmeval library and standard bias formulas. For Model Monitor, it points to open-source monitoring solutions built on MLflow and Evidently AI, with CloudWatch and QuickSight for dashboards. AWS’s own wording for the group, from its Services in Maintenance reference: “Customers can’t on-board to services and features in the maintenance stage. Customers already using these services and features can continue to do so. AWS will continue to operate and support these services and features but won’t enhance or add functionality to them.”

Two SageMaker items go further than maintenance mode. Ground Truth Plus, the managed labeling service, reached full end of support on 30 June 2026. SageMaker Profiler is entering sunset: closed to new customers on 30 July 2026, with full end of support on 30 June 2027.

What “closed to new customers” really means

AWS’s wording is careful, and it is worth translating into what it means in practice. “Closed to new customers” is scoped to your AWS account. If your account was not already using the service before the cutoff date, you cannot start using it at all: the create actions are blocked. Amazon Bedrock Agents Classic is explicit about the mechanism. It allowlists only accounts with usage in the past twelve months and returns an access-denied error to everyone else.

So two things are true at once. For a team already running the service, it is “still fully supported.” For everyone else, a new account, a new project, or a colleague who never touched it, the service is effectively gone. That is why, for any green-field decision today, every service in this wave is off the table: you either already depend on it, or you can no longer choose it. In that practical sense these services are deprecated for new work, even though AWS avoids the word. “Closed to new customers” is the polite phrasing for “unavailable unless you were already here.”

Why it matters for builders

If your architecture rests on any of these, the clock is a planning input, not an emergency. The practical steps: check whether you have workloads on Kendra, Bedrock Agents Classic, or Q Business; note that maintenance mode still runs, so nothing breaks on the cutoff date, but new features stop and new accounts are locked out; and schedule migration to the named successor before you build anything new on the old service. Teams still on Amazon Q Developer’s IDE plugins have until April 2027 and a clear path to Kiro.

The wider lesson is the one this wiki keeps returning to: on a managed cloud, the platform’s lifecycle is part of your architecture. A service you picked for a good reason can enter maintenance mode on the vendor’s schedule, not yours. The defence is the same as for model deprecations: know which of your dependencies are load-bearing, keep an eye on lifecycle notices, and avoid designs that are expensive to unwind. See service lifecycle and deprecation for how to plan around it, and preparing for AI provider restrictions for the same discipline applied to models.

Update: two more AWS retirements since this wave

This tracker keeps growing. Two more AWS deprecations, unrelated to the original AI-stack consolidation described above, land in the same late-September window:

ServiceWhat changedKey dateAWS points you to
AWS App MeshFull end of support; console and resources become inaccessible30 September 2026Amazon ECS Service Connect (for ECS) or Amazon VPC Lattice (for EKS and cross-account networking)
Amazon Nova Reel v1:1 (Bedrock model)Model reaches end of life; requests fail after this date30 September 2026A newer, Active Nova Reel model version on Bedrock

AWS App Mesh is a true shutdown like AWS Proton above, not another “maintenance mode” step. Existing customers can keep creating App Mesh resources right up to the cutoff, and AWS keeps shipping security and availability updates until then, but after 30 September 2026 the App Mesh console and all App Mesh resources become inaccessible. AWS’s migration guidance splits by workload: Amazon ECS users are pointed to ECS Service Connect, while Amazon EKS and cross-account users are pointed to Amazon VPC Lattice.

Nova Reel v1:1 is a model retirement, not a service shutdown, so it runs on Bedrock’s separate model lifecycle rather than the “maintenance mode” track above: it entered Legacy status on 30 March 2026 and reaches End-of-Life on 30 September 2026, after which requests to that specific model ID (amazon.nova-reel-v1:1, offered in us-east-1) fail across all regions, with no automatic migration. This is narrower than it sounds: Bedrock’s model-lifecycle table also lists amazon.nova-reel-v1:0 and Amazon Nova Canvas reaching the same 30 September 2026 EOL date, so check which specific Nova model IDs your application is pinned to.

Beyond AWS: two more deprecations landing the same week

Two non-AWS deprecations fall in the same narrow window and are worth flagging for anyone building across clouds, even though they sit outside this article’s AWS focus.

ServiceVendorWhat changedKey dateMigrate to
Gemini Omni Flash Preview (gemini-omni-flash-preview)GooglePreview model shuts down30 September 2026gemini-omni-1.1-flash
Azure AI PersonalizerMicrosoftFull service retirement1 October 2026Open-source microsoft/learning-loop

gemini-omni-flash-preview shipped in public preview on 30 June 2026 and, per Google’s own Gemini API deprecations page, shuts down three months later on 30 September 2026 in favor of the now-generally-available gemini-omni-1.1-flash. That is a short preview lifecycle even by Gemini’s standards, and a reminder that “preview” model IDs from any provider are not meant for production pinning.

Azure AI Personalizer, the reinforcement-learning content-selection service covered on this wiki at Azure Personalizer , has been closed to new resources since 20 September 2023 and is now scheduled for full retirement. Microsoft’s own documentation is genuinely inconsistent here, and it is worth being precise about which sources say what. The Personalizer product docs (now archived, as of 1 September 2026, into Microsoft’s “previous versions” tier) state “the Personalizer service is being retired on the 25th of August, 2026.” The Microsoft Lifecycle page for Personalizer gives a third date in its own structured Support Dates table — 26 August 2026 — while separately linking out to a Microsoft announcement titled “AI services: Personalizer will be retired on 1 October 2026” (that announcement page no longer resolves). Weighing against the August dates: Azure Advisor’s Service Upgrade and Retirement reference, a comprehensive, actively maintained tracker of every Azure retirement, lists Azure Personalizer retiring 1 October 2026, grouped in the same batch as Azure Anomaly Detector and Azure Metrics Advisor — which matches independent trackers and is the most recently confirmed source we found. That is the date to plan against, though given the conflicting August dates elsewhere in Microsoft’s own docs, confirm against the Azure Advisor recommendation for your own subscription before treating either date as final. Microsoft recommends migrating to the open-source microsoft/learning-loop project.

For the general pattern behind all of these moves, not just AWS, see service lifecycle and deprecation .

Sources

Further reading