Amazon Bedrock Agents is closed to new customers. What should answer questions about your S3 documents now?

What actually changed

Bedrock Agents launched in November 2023 as the managed way to build an agent: pick a model, write instructions, attach action groups and a knowledge base, and let Bedrock run the orchestration loop. Since July 30, 2026 that service is Agents Classic. Accounts with Bedrock Agents activity in the previous 12 months are allowlisted and keep working. Everyone else gets an AccessDeniedException from CreateAgent and InvokeInlineAgent, and AWS states there is no exception process.

The rest of Bedrock is not affected. Model inference, Knowledge Bases and Guardrails continue, and new models keep arriving for them. The Agents Classic model catalogue, by contrast, is frozen at the July 30 list, no new features are planned, and there is no end-of-life date. Existing agents, their APIs, CloudFormation resources and IAM action prefixes carry on unchanged.

Amazon Bedrock AgentCore for document agents on S3

• Closed to new accounts: CreateAgent and InvokeInlineAgent return 403 without prior usage.
• Still open: Bedrock models, Knowledge Bases, Guardrails, and every existing Agents Classic agent.
• Frozen: the model list available inside Agents Classic.
• No deadline: AWS has set no end-of-life date, but recommends moving to AgentCore.

What AWS recommends instead

AWS points new agent work at Amazon Bedrock AgentCore, and specifically at its managed harness. The harness is the closest thing to the old experience: you declare a model, a system prompt and a set of tools, and AgentCore runs the loop, the compute, memory, identity and tracing. For anything the harness cannot express, you deploy a code-defined agent on AgentCore Runtime using Strands, LangChain, the OpenAI Agents SDK, the Claude Agent SDK or your own code.

The feature map matters if you had an Agents Classic design in your head. Action groups become MCP tools behind AgentCore Gateway, wrapping REST APIs, Lambda functions or code-level tools. A knowledge base attaches through Gateway or a retrieval tool rather than directly on the agent. Return of control becomes an inline function tool: the agent pauses and hands a tool call back to your client code. Stage-specific prompt overrides are not replicated, and routing-style multi-agent collaboration still needs custom code.

The commands below are the minimal harness setup from AWS's migration guide, with the model and prompt swapped for a document assistant:

agentcore create --name contracts-assistant
agentcore add harness \
  --name contracts-assistant \
  --model-id us.anthropic.claude-sonnet-4-6-20250514-v1:0 \
  --system-prompt "You answer questions about signed supplier contracts and cite the clause." \
  --tools code-interpreter
agentcore deploy

• Harness: config-based, managed loop, closest analog to Agents Classic.
• Code-defined agents: any framework, full control of the loop, more work.
• An AWS-published skill in the agent toolkit can import an existing Agents Classic configuration as a starting point.

What a document agent on AgentCore still has to solve

Neither Agents Classic nor AgentCore ships a document connector for S3. On AgentCore your agent reads files through the AWS SDK under its execution role, or through a Gateway tool, usually a Lambda that lists and fetches objects. You write the policy that grants s3:GetObject, and whatever scope you give the role is what the agent can read for every user. Per-user visibility means passing identity through AgentCore Identity and enforcing it in your tool code.

The model limits are the same as everywhere else on Bedrock. A Converse request takes up to five documents of 4.5 MB each, so an agent working a folder of contracts reads a few at a time, and every page it reads is input tokens on your bill. AgentCore's own meters, Runtime vCPU-hours and GB-hours plus Gateway invocations, come on top. AWS notes that the harness is more token-efficient than Agents Classic's internal prompts, so inference cost may be comparable or lower than the old service.

Ask questions about S3 documents with Claude on Amazon BedrockAmazon Bedrock pricing for document Q&A

• You still write the S3 access policy and any per-user rules.
• Five documents of 4.5 MB per model call, billed per token.
• AgentCore compute and Gateway charges are separate from model inference.

Amazon Quick: the index-based route

Amazon Q Business, the other product people searched for, was closed to new customers earlier in 2026, and AWS points those teams to Amazon Quick. Quick is a finished workspace rather than a toolkit: it crawls connected sources including S3 into an index, and people chat and search across it. It can also call MCP servers and AgentCore Gateways, so an agent you build on AgentCore can appear inside Quick.

An index answers a different question from an agent. "Which of our supplier contracts auto-renew this quarter?" needs an index across thousands of files, and that index costs money every hour whether anyone asks or not. It also needs per-document access rules carried into the index, which is the step migrations most often get wrong.

Amazon Bedrock Knowledge Bases on S3, or just ask the document?

• Choose Quick when the question spans many files and the answer is a search.
• Budget for the index to run continuously.
• Carry document permissions into the index on day one.

BucketDesk Document AI: ask the file in front of you

Most document questions in a business are about one file: what is the notice period in this contract, what did this invoice charge for, what changed between these two versions of the policy. For that, no agent and no index is required. BucketDesk Document AI lets you open a file in a connected bucket, ask in plain language, and get an answer with a citation to the supporting passage, generated by a Bedrock model running in your own AWS account.

Because it runs in your account, your Bedrock settings apply: your data retention mode, your model choice, your bill per request. It reads with the access the person already has in BucketDesk, so there is no execution role to scope or ACL file to maintain, and it keeps no file content after the session. It does not run multi-step workflows or act on other systems; if an invoice has to be matched against a purchase order and filed, that is a job for an agent on AgentCore.

Which Amazon Bedrock model should answer questions about your documents?Document AI on the features pagePlans and the Ultra trial

• No agent code, index or execution role to maintain.
• Citations to the passage the answer came from.
• Runs on Bedrock in your account; your retention mode and model settings apply.
• Does not take actions in other systems.

Which one, in one pass

Start from what the system has to do with the document, not from which AWS product is newest.

• Answer a question about the file someone has open: BucketDesk Document AI.
• Search and chat across thousands of files with permissions enforced: Amazon Quick, with the index cost accepted.
• Read documents and then act, such as comparing invoices to a purchase order system and filing a result: the AgentCore harness, or a code-defined agent if you need custom orchestration.
• Keep an existing Agents Classic agent: it works, but plan the AgentCore migration since the model list is frozen.
• Starting in a brand-new AWS account: Agents Classic is not an option at all.