AWS Closes Bedrock Agents: AgentCore Migration 2026

Amazon Bedrock Agents is now Amazon Bedrock Agents Classic, and as of July 30, 2026 it is closed to new customers.¹ Existing agents keep running with no end-of-life date, but the model catalog is frozen and no new features are planned. AWS recommends migrating to Amazon Bedrock AgentCore.

TL;DR

The product AWS launched as Agents for Amazon Bedrock at re:Invent 2023 is in maintenance mode two years and eight months after general availability.²¹

Nothing breaks today. Every existing API — InvokeAgent, UpdateAgent, action group and knowledge base APIs — still works.¹

What changed is the door. Accounts with no Bedrock Agents activity in the previous 12 months now get an AccessDeniedException (HTTP 403) on CreateAgent and InvokeInlineAgent, and AWS states plainly that there is no exception process.¹

The quieter cost is the frozen model catalog. Any foundation model AWS adds to Bedrock after July 30, 2026 is reachable from AgentCore but not from Agents Classic.¹

Bedrock Agents was not singled out. It arrived in a batch of 20 services and features moved to maintenance in a single June 30, 2026 announcement, alongside Amazon Kendra and Amazon Q Business.³

What You'll Learn

• What "Agents Classic" actually means for a running workload
• Whether your AWS account is allowlisted, and how to tell
• The four capability gaps AWS documents in its own migration table
• What AWS says migration costs — and why the number circulating in blog coverage is wrong
• How this fits the wider first-generation AI service cleanup at AWS

What "maintenance mode" means here

Maintenance mode at AWS is not deprecation. It is a closed door with the lights left on.

AWS's own wording: "Amazon Bedrock Agents (launched November 2023) is now Amazon Bedrock Agents Classic and will no longer be open to new customers starting on July 30, 2026."¹

The FAQ is explicit that existing agents "continue to function normally," that there is no migration deadline, and that Bedrock Agents Classic "remains available to existing customers in maintenance mode with no planned end-of-life date."¹

Bedrock itself is untouched. AWS states that Amazon Bedrock — model inference, Knowledge Bases, and Guardrails — "continues to receive new models," and that the restriction "applies only to the Bedrock Agents Classic orchestration layer."¹

The name change is cosmetic at the API level. Existing namespaces (bedrock-agent), SDK clients, CloudFormation resource types, and IAM action prefixes are unchanged, and existing IaC templates keep working for allowlisted accounts.¹

There is also no bill attached to Classic. AWS confirms "there is no charge for Bedrock Agents Classic itself" — you pay for model inference and associated resources, as before.¹

Are you allowlisted?

This is the one question worth answering today.

AWS determines the allowlist automatically: accounts with Bedrock Agents activity in the past 12 months are allowlisted and unaffected.¹

It applies per account, not per organisation. If you run a multi-account setup, only the accounts with prior Bedrock Agents usage are allowlisted.¹

That is the sharp edge. A team whose production account is allowlisted can still hit a 403 the first time it spins up a fresh sandbox, staging, or per-developer account and runs the same CloudFormation template that worked last quarter.

The error is specific: AccessDeniedException, HTTP 403, with the message "Bedrock Agents is in Maintenance Mode. New agent creation is not available for accounts without prior service usage."¹

Inline agents follow the same rule. Accounts with InvokeInlineAgent usage in the past 12 months can continue; AWS says migration guidance for inline agents is "forthcoming."¹

The four gaps AWS documents itself

The most useful thing in the migration page is not the recommendation. It is the comparison table, where AWS lists what does not carry over.

Read against the marketing, it is unusually candid. Four rows describe a downgrade rather than a swap.

There is real irony in the second row. Multi-agent collaboration was one of the headline additions to Bedrock Agents, reaching general availability on March 10, 2025 after a re:Invent 2024 preview⁴ — and it is the capability the recommended successor handles least well.

The pattern across all four rows is the same: things that were declarative configuration in Classic become code you own in AgentCore. That is defensible as an architecture choice. It is not a like-for-like migration, and AWS does not claim it is.

What AWS actually says migration costs

Secondary coverage has circulated a number and attributed it to AWS: "AWS estimates the migration timeline at two to four weeks for most workloads."⁵ That figure does not appear in AWS's migration documentation.

AWS's own estimate is smaller and more conditional: "For straightforward agents (model + action groups + knowledge base), CLI import or harness setup takes hours."¹

The caveat carries the weight: "Complex agents with custom orchestrators or multi-agent collaboration will require more significant code work."¹

So the honest read is a split population, not an average. Config-shaped agents move in an afternoon. Agents that leaned on prompt-stage overrides or routing-mode collaboration are a rewrite, and no published estimate covers them.

The official tooling is an agent skill rather than a converter. AWS's migration page points to an amazon-bedrock skill in the agent toolkit for AWS that inspects the existing agent, checks migration eligibility, maps each component to a harness equivalent, and drives the AgentCore CLI — and AWS notes it "never modifies your source agent."¹⁶

Separately, agentcore import-agent translates a Bedrock Agent's configuration into runnable code, with the target platform documented as strands (1.0.x) or langchain (0.3.x) + langgraph (0.5.x).⁷ It writes out the converted agent, a requirements.txt, an .agentcore-config.yaml, and a generated README.⁷

Worth checking which repository you are following: the bedrock-agentcore-starter-toolkit that documents import-agent is now labelled legacy, with new projects pointed at the AgentCore CLI.⁷

Where AgentCore actually is

The replacement is not new. AgentCore went to preview on July 16, 2025⁸ and reached general availability on October 13, 2025 in nine AWS Regions.⁹

The piece that makes it a genuine Classic substitute is newer. The managed agent harness — the config-based path where you declare model, tools, and instructions instead of writing the loop — only reached general availability on June 17, 2026.¹⁰

That is six weeks before the door closed on Classic. AWS gave its config-first replacement about a month and a half of GA maturity before restricting the thing it replaces.

The timing is not an AWS quirk. Microsoft's Agent Framework harness reached stable release on July 22, 2026, five weeks after AWS's, with a near-identical config-over-code framing.

Region coverage still lags the platform. AgentCore's documentation lists 20 regions, and the harness is available in 16 of them — not yet Europe (Milan), Europe (Spain), Asia Pacific (Malaysia), or Asia Pacific (Thailand).¹¹ AWS acknowledges the mismatch directly, telling customers in unsupported regions to keep running Classic there while migrating new development elsewhere.¹

On the other side of the ledger, AgentCore is a real expansion. It supports the full Bedrock catalog plus OpenAI, Gemini, and any OpenAI-compatible endpoint, with provider switching mid-session and no redeploy.¹

Runtime has also moved since its own GA. It launched with eight-hour execution windows, complete session isolation, and Agent-to-Agent (A2A) protocol support.⁹ An Instances compute type added in August 2026 runs agents on AWS-managed EC2 in your own account and supports persistent sessions of up to 14 days.¹²

A single CLI command exports the harness to Strands-based code, with Claude Agent SDK listed as a coming export target.¹⁰

Pricing is the change most likely to surprise a finance team, and it lands in a year when per-session agent spend caps became a mainstream ask.

Billing moves from free-with-inference to consumption-based across runtime, memory, and gateway. AWS notes there is no separate harness orchestration charge, and argues the harness prompts are more token-efficient than Classic's internal ones.¹

That last claim is AWS's own. The underlying comparison is not published, and no independent verification of it had surfaced as of August 17, 2026.

For a wider view of how AWS, Google, and Alibaba are each packaging the agent loop as billable infrastructure, see our earlier breakdown of managed agent runtimes across the three clouds.

The batch this arrived in

Bedrock Agents did not get a dedicated obituary. It appeared as one bullet in an AWS Service Availability Updates post on June 30, 2026.³

That post moves 20 entries to maintenance: 11 named services and features — including Amazon Cognito Sync, AWS Directory Service Simple AD, and AWS Systems Manager Application Manager — plus 9 SageMaker AI features such as Clarify, Debugger, Ground Truth, Model Monitor, and Studio Lab.³

Two neighbours in that list matter for agent builders, because together with Bedrock Agents they covered the retrieval, assistant, and orchestration layers many teams built on.

Amazon Kendra entered maintenance mode effective June 30, 2026 and closed to new customers on July 30, with Bedrock Managed Knowledge Bases as the recommended path.¹³ The connector gap there is stark: Kendra supports 32-plus native connectors, BMKB supports seven.¹³

Amazon Q Business is likewise closed to new customers, with Amazon Quick as the successor and a Bring Your Own Index path that carries the existing index across.¹⁴

These are not a single cohort, which makes the pattern more telling. Kendra reached general availability on May 11, 2020,¹⁵ Bedrock Agents on November 28, 2023,² and Q Business on April 30, 2024.¹⁶ Their runs to July 30, 2026 were roughly six years, two years eight months, and two years three months respectively — closed to new customers in one post.

The common thread is not failure. It is that all three were built around the same assumption — a managed service per task — while the 2026 generation is built around a runtime that hosts whatever framework you bring. Kendra to a knowledge base primitive, Q Business to an assistant platform, Bedrock Agents to a hosting layer. Same direction, three times.

Teams that read the July 2026 news as "AWS killed Bedrock Agents" got the framing wrong. Teams that read it as "nothing changes" will find out at the next new-account bootstrap.

What to do this week

1. Enumerate accounts, not workloads. Run a Bedrock Agents API call in every account you provision into. Allowlisting is per account and invisible until something 403s.
2. Check your region against the harness table. If you are in Milan, Spain, Malaysia, or Thailand, the config-based path is not there yet.¹¹
3. Classify each agent as config-shaped or code-shaped. Model plus action groups plus knowledge base migrates in hours. Prompt-stage overrides or routing-mode multi-agent do not.¹
4. Decide about the frozen catalog, not the deadline. There is no deadline. The forcing function is the next model you want that Classic cannot reach.¹
5. Test with import-agent before committing. It generates the converted agent code, a requirements.txt, and an .agentcore-config.yaml you can read before deploying anything.⁷

The bottom line

The headline number is 32 months — Bedrock Agents went from general availability on November 28, 2023 to closed-to-new-customers on July 30, 2026.²¹

The number that should worry a platform team is different: nine and a half months. That is how long AgentCore had been generally available when its predecessor's door closed, and six weeks is how long the harness — what AWS calls "the closest analog to the Bedrock Agents managed experience" — had been GA.¹⁹¹⁰

Nothing here is an outage. But the pace tells you what to assume about managed agent abstractions generally: the loop is the product now, and the vendor reserves the right to re-architect it faster than you can amortise a migration.

Budget for that, not for a deadline that does not exist.