AWS Retired Bedrock Agents Classic: The AgentCore Shift and the Lock-In Lesson Every Enterprise Should Learn

If your company has spent the last two years building AI agents on Amazon Bedrock, you got some unwelcome news on July 30, 2026: the service you built on has a new name, a frozen feature set, and a large “please migrate” sign hanging over it. AWS didn’t announce this with a keynote or a splashy blog post. It quietly renamed Amazon Bedrock Agents to “Bedrock Agents Classic,” closed it to new customers, and put it into maintenance mode — pointing everyone toward its replacement, Bedrock AgentCore.

Taken on its own, this is a routine cloud-vendor product transition. But it’s a useful data point in a bigger story: every major cloud provider is re-architecting its agent platform out from under the enterprises that bet on the first version, and each rebuild locks customers in a little tighter to that provider’s identity, memory, and governance stack. The Bedrock Agents shutdown is worth understanding in detail — not because you’re necessarily on AWS, but because the same pattern is coming for whichever platform you did pick.

What Actually Changed on July 30

Amazon Bedrock Agents launched in November 2023 as a fairly simple single-agent framework: you defined an agent, gave it action groups (tool calls wrapped in OpenAPI schemas), attached a knowledge base for retrieval, and let a managed orchestration loop run the show. As of July 30, 2026, that product — now called Bedrock Agents Classic — stopped accepting new customers. Per AWS’s own migration documentation, the practical effects are narrow but real:

• Accounts with no Bedrock Agents activity in the prior 12 months lose access to the CreateAgent and InvokeInlineAgent APIs outright — they get an AccessDeniedException if they try.
• Existing agents keep running. InvokeAgent, UpdateAgent, ListAgents, knowledge base APIs, and alias APIs all remain available to allowlisted accounts.
• The model catalog available inside Bedrock Agents Classic is frozen as of that date. Any model released after July 30, 2026 will only be available through the new platform.
• There’s no announced end-of-life date and no forced migration deadline — but no new features are coming to the Classic product, either.

In other words: nobody’s agents broke overnight. But anyone building new agent workloads on AWS is now, by policy, building on Bedrock AgentCore instead — a very different platform under the hood.

Why AWS Rebuilt Rather Than Patched

AgentCore isn’t Bedrock Agents with a new logo. It went generally available back in October 2025 as a full re-architecture organized around five separate managed services: Runtime (serverless execution with session isolation and multi-hour windows), Memory (configurable short- and long-term memory strategies), Gateway (turns REST APIs, Lambda functions, and MCP servers into agent-callable tools), Identity (OAuth-based, identity-aware authorization with token vault storage), and Observability (OpenTelemetry-compatible tracing that plugs into CloudWatch, Datadog, or Dynatrace).

The reason for the rebuild is basically the same reason every large agent platform has been rebuilt in the past year: the first generation of agent frameworks assumed one agent, calling a handful of tools, inside one request-response loop. What enterprises are actually deploying now looks nothing like that — it’s fleets of agents that hand off tasks to each other, persist memory across sessions, need per-agent identity for audit trails, and route tool calls through a shared, governed gateway rather than each agent hardcoding its own integrations. That’s the same governance gap we covered in our look at enterprise agent fleets: once you have more than a couple of agents in production, ad hoc tooling and access control stop scaling. AgentCore’s Gateway component leans specifically on the Model Context Protocol to expose tools in a standard, framework-agnostic way — which matters for the lock-in question below.

AWS also built two on-ramps rather than one. The managed harness is a declarative, config-based experience close to what Bedrock Agents Classic customers already know — you declare a model, tools, and a system prompt, and AgentCore handles compute, memory, and observability for you. For teams that need custom orchestration, multi-agent supervisor patterns, or a specific framework (Strands, LangChain, the OpenAI Agents SDK, or the Claude Agent SDK), there’s also a code-defined path that still runs on AgentCore’s managed infrastructure but hands you the orchestration loop directly.

The Migration, in Practice

Early evidence suggests the actual migration effort is smaller than the vendor-transition anxiety implies. One developer who moved a production AWS-billing-analysis agent off Classic documented the process: roughly 90 minutes of hands-on work, a monthly cost increase of about 15 cents, and — after running both versions in parallel for a week — functional parity on six of seven test days, with the new build actually outperforming the old one on one edge case. Tools that used to be separate Lambda functions wrapped in OpenAPI schemas became plain Python functions decorated with @tool, and local testing went from minutes-per-iteration to seconds-per-iteration.

AWS also shipped an automated path: a skill in its agent toolkit for AWS that inspects an existing Bedrock Agent, checks migration eligibility, maps each component to its AgentCore equivalent, and scaffolds the new deployment — pausing for human approval before anything deploys. That’s meaningfully lower-friction than most vendor platform transitions offer, and worth noting as a positive signal even if it doesn’t remove the underlying lesson.

The Pattern Isn’t Unique to AWS

This is the part that matters even if you have zero AWS workloads. Google spent 2026 doing something similar and more aggressive: Vertex AI, as a standalone product, no longer exists in the Google Cloud console. It’s been fully absorbed into the Gemini Enterprise Agent Platform, reorganized around four pillars — Build, Scale, Govern, and Optimize — that bundle agent authoring, memory, a governed agent registry, and evaluation into one product. Microsoft has taken the opposite tactic of deepening rather than renaming: Azure AI Foundry ties agent identity directly to Entra ID, so agents authenticate through the same conditional-access policies as human employees — a genuinely useful governance feature, and also a very effective way to make switching cloud providers later a much bigger project than it would otherwise be.

An analysis comparing the three platforms put it bluntly: “your existing cloud infrastructure will likely determine your agent platform more than any feature comparison” — and switching costs compound as you build more agents on top of any one of them. None of the three vendors are doing anything wrong here; deeper identity and memory integration is genuinely what production agents need. But it’s also, mechanically, lock-in, and it’s arriving faster than most enterprise architecture reviews are built to catch.

What This Means for Your Agent Roadmap

A few concrete takeaways, regardless of which cloud you’re on:

1. Treat your tool layer as the abstraction boundary, not the agent framework. AgentCore’s Gateway exposing tools as MCP servers is the detail to copy even if you never touch AWS: if your agents call tools through a protocol-standard interface rather than vendor-specific SDK calls, swapping the orchestration layer underneath becomes a rewrite of glue code, not a rewrite of your business logic.
2. Budget migration time into any “managed agent platform” decision, not just the build. The Bedrock migration data point — 90 minutes of real work for a moderately complex agent — is reassuring, but it assumes you built cleanly in the first place. Agents with custom orchestrators, deep prompt-override chains, or hardcoded action-group schemas will cost more to move.
3. Separate “no deadline” from “no urgency.” AWS is explicit that there’s no forced migration date for Bedrock Agents Classic. That’s true, and it’s also exactly the situation that lets technical debt quietly compound for two years until a security review, an audit, or a genuinely useful new model release forces the migration at the worst possible time.
4. Governance and identity integration are now the actual differentiator between agent platforms — not raw model quality, which is converging across providers. If you’re evaluating a platform for the first time, weight Identity and Observability features as heavily as Runtime capability, because that’s the layer everyone is now building lock-in around, whether intentionally or not. This is the same underlying risk our guide to scaling AI enterprise-wide flags: production agents create governance surface area that pilots never do, and vendor platform choice is now part of that surface, not separate from it.

None of this means avoid managed agent platforms — building your own Runtime, Memory, Gateway, and Identity layer from scratch is a much bigger commitment for almost every team than accepting a vendor’s version. It means going in with your eyes open about which parts of the stack are genuinely portable and which parts you’re choosing to marry.

Frequently Asked Questions

Do I need to migrate off Bedrock Agents Classic right now?

No. AWS has confirmed there is no forced migration deadline, and existing agents continue to function normally with full API support. However, no new models or features will be added to Bedrock Agents Classic going forward, so new development should target AgentCore.

Will my existing Bedrock Agents stop working after July 30, 2026?

No. The change only affects new agent creation for accounts without prior Bedrock Agents usage in the past 12 months. Existing agents, knowledge bases, and guardrails continue to operate as before.

Is Amazon Bedrock AgentCore more expensive than Bedrock Agents Classic?

Not necessarily. AWS states there’s no separate charge for AgentCore’s harness orchestration, and because its managed harness is more token-efficient than Bedrock Agents Classic’s internal prompting, some customers may see comparable or lower model inference costs alongside AgentCore’s own consumption-based fees for runtime, memory, and gateway usage.

How is this different from what Google and Microsoft are doing?

Google folded Vertex AI entirely into the Gemini Enterprise Agent Platform, removing the standalone product from its console. Microsoft has kept Azure AI Foundry as a distinct product but tied agent identity directly into Entra ID. All three moves push customers toward deeper platform-specific integration, just through different mechanisms.

What’s the single best way to reduce lock-in risk when adopting any of these platforms?

Standardize your agent’s tool integrations on the Model Context Protocol wherever the platform supports it. Since MCP is an open, framework-agnostic standard rather than a vendor SDK, tools built against it are far cheaper to carry over if you later change orchestration platforms or cloud providers.

Can I use non-AWS models inside Bedrock AgentCore?

Yes. AWS documentation confirms AgentCore supports the full Bedrock model catalog plus additional providers, including OpenAI- and Gemini-compatible endpoints, and lets you switch model providers mid-session without redeploying the agent.

Sources

• Amazon Bedrock Agents Classic maintenance mode - AWS Documentation - AWS’s official migration guide and FAQ covering the July 30, 2026 policy change.
• Amazon Bedrock AgentCore is now generally available - AWS - AWS’s official GA announcement for AgentCore, October 2025.
• Vertex AI Is Now Gemini Enterprise Agent Platform: What Changed in 2026 - GCP Study Hub - Overview of Google’s rebrand of Vertex AI into an agent-centric platform.
• Microsoft Foundry vs AWS Bedrock AgentCore vs Google Vertex AI: The 2026 Enterprise Agent Platform Decision Guide - AgentMarketCap - Comparative analysis of lock-in factors across the three major cloud agent platforms.
• My Robot Accountant Got an Eviction Notice. So We Moved: Migrating from Bedrock Agents Classic to AgentCore - DEV Community - A developer’s first-hand account of migrating a production agent, including time and cost figures.
• The Enterprise AI Platform Wars in 2026: AWS Bedrock vs Azure AI Foundry vs Google Vertex AI - Broader industry context on the hyperscaler agent platform competition.
• Agent-as-a-Service: Comparing Claude Managed Agents and Amazon Bedrock AgentCore - DEV Community - Technical comparison of managed agent infrastructure approaches.