Amazon Bedrock agents built to pass security review and keep running after launch.
On 30 July 2026 AWS stopped new customers starting on Bedrock Agents, now called Bedrock Agents Classic, and pointed teams to AgentCore. We build agents and retrieval on Bedrock and plan identity, guardrails, evaluations and model lifecycle from sprint one.
Bedrock Agents Classic, closed to new customers 30 Jul 2026
Evaluations and Policy
Evaluations GA 31 Mar 2026, Policy GA 3 Mar 2026
release watch
AgentCore generally available13 Oct 2025
AgentCore Evaluations GA31 Mar 2026
connects with
in short·15-second read
We build and run production AI on Amazon Bedrock.
Agents go on Amazon Bedrock AgentCore, generally available since October 2025, and reach internal APIs through AgentCore Gateway and Identity. We set up Knowledge Bases and Guardrails and add AgentCore Evaluations and Policy. Model upgrades get planned against Bedrock's Active, Legacy and EOL lifecycle, so a retirement notice doesn't halt production.
you probably need this if
A Bedrock Agents Classic prototype works. Nobody has planned how it moves to AgentCore.
Security asked how the agent authenticates to internal APIs. Nobody had an answer.
Retrieval quality was judged by a demo, not a test set of real questions.
13 Oct 2025AgentCore generally available Runtime, Memory, Gateway, Identity and Observability GA in nine Regions. source ↗
31 Mar 2026AgentCore Evaluations GA Thirteen built-in evaluators. AgentCore Policy became GA on 3 March 2026. source ↗
30 Jul 2026Bedrock Agents renamed Agents Classic Renamed Amazon Bedrock Agents Classic, and closed to new customers on 30 July 2026. Existing customers can carry on using it. source ↗
07 Sep 2026New model lifecycle policy Bedrock models move from Active to Legacy to EOL. Models launched on or after 7 September 2026 follow a new policy. source ↗
today · Oct 22, 2026
the story · amazon bedrock estate
From the first call to a setup your own team can run.
Here's what usually prompts a call about your Amazon Bedrock estate, what we do about it and in which order, and where you end up.
01 · where it hurts02 · what we change03 · how it runs04 · where you end up
incident board · amazon bedrock estate3 open
openA Bedrock Agents Classic prototype works. Nobody has planned how it moves to AgentCore.
recurringSecurity asked how the agent authenticates to internal APIs. Nobody had an answer.
blockedRetrieval quality was judged by a demo, not a test set of real questions.
exposure
Agents on AgentCoreAgent identity and accessKnowledge Bases and retrievalGuardrails, Policy and EvaluationsMigration off Bedrock Agents…Model lifecycle and cost
startrehearsehandover
Use case and data review
Identity and access design
Build and evaluate
Launch and handover
we are here
ready to hand over
It's on the platform AWS is investing in.
The model changes before EOL with a measured comparison, and the team has a test set for…
Use case brief with systems, data owners and success criteria
Evaluation set from real past cases, run in CI
chapter 01 · where it hurts
Why teams reach out
A Bedrock Agents Classic prototype works. Nobody has planned how it moves to AgentCore.
Security asked how the agent authenticates to internal APIs. Nobody had an answer.
Retrieval quality was judged by a demo, not a test set of real questions.
chapter 02 · what we change
How we solve it
Agents on AgentCore
Agent identity and access
Knowledge Bases and retrieval
Guardrails, Policy and Evaluations
Migration off Bedrock Agents Classic
Model lifecycle and cost
chapter 03 · how it runs
Our approach
Use case and data review
Identity and access design
Build and evaluate
Launch and handover
chapter 04 · where you end up
What changes
It's on the platform AWS is investing in.
The model changes before EOL with a measured comparison, and the team has a test set for the next retirement notice.
Use case brief with systems, data owners and success criteria
Evaluation set from real past cases, run in CI
what we do
Our Amazon Bedrock work
6 kinds of Amazon Bedrock work, each run by engineers who have done it on live tenants.
Agents deployed on AgentCore Runtime, with Memory for session context, Gateway for tool access to internal APIs and Observability traces wired into your monitoring.
AgentCore Identity configured so agents act with scoped credentials on behalf of users or as workloads, with Entra ID or Okta as the identity provider.
Bedrock Knowledge Bases over approved document sources, with chunking, metadata filters and agentic retrieval tested against real questions.
Bedrock Guardrails for content and contextual grounding, AgentCore Policy for what agents may do, and AgentCore Evaluations run before every release.
Existing Bedrock Agents Classic action groups, prompts and knowledge base links rebuilt on AgentCore, with behaviour compared on the same test set.
Model choice per use case, cross-Region inference profiles, and a plan for each model's Legacy and EOL dates.
capability 01 / 06in production
Agents on AgentCore
Agents deployed on AgentCore Runtime, with Memory for session context, Gateway for tool access to internal APIs and Observability traces wired into your monitoring.
AgentCore Identity configured so agents act with scoped credentials on behalf of users or as workloads, with Entra ID or Okta as the identity provider.
We pieced these together from Amazon Bedrock situations that come up a lot. They're patterns. You won't find a specific client in them.
case file 01
A Bedrock Agents Classic prototype, three action groups, on its way to production
composite · not a client story
Beforethe situation
Someone built a claims triage agent on Bedrock Agents, with three Lambda action groups and one knowledge base. That service is now Agents Classic, closed to new customers, and security is asking for scoped credentials per tool call.
What we dothe work
We build an evaluation set out of past claims, then rebuild the agent on AgentCore Runtime and put the Lambda functions behind AgentCore Gateway, with credentials coming from AgentCore Identity. Both versions run the same tests, and we compare.
Afterwhere you end up
It's on the platform AWS is investing in. You can trace each tool call to an identity, and security now has evidence to sign off on, where before it had a diagram.
case file 02
A model heading to Legacy under a live customer-service assistant
composite · not a client story
Beforethe situation
A customer-service assistant runs on a model that has moved to Legacy. The prompts were tuned by hand and there is no record of how good the answers were at launch.
What we dothe work
We build an evaluation set from transcripts, score the current model to set a baseline, then test two candidate replacements with Bedrock evaluations and switch behind a feature flag.
Afterwhere you end up
The model changes before EOL with a measured comparison, and the team has a test set for the next retirement notice.
We don't start a step until the last one has produced something you can check.
step 01
Use case and data review
We pick a single workflow, list what systems and documents the agent will need, check who owns each, then write the first evaluation set out of real past cases.
step 02
Identity and access design
We design AgentCore Identity, Gateway targets and IAM roles so each tool call carries scoped credentials, and go through the design with your security team before we build.
step 03
Build and evaluate
We put the agent on AgentCore Runtime, with Knowledge Bases and Guardrails. AgentCore Evaluations and Policy go into the pipeline, and you can see the traces in your own monitoring.
step 04
Launch and handover
Real users get the agent in stages. We keep a register of each model's Legacy and EOL dates, and the team that owns it next gets the runbooks.
ai on amazon bedrock
Where AI fits on Amazon Bedrock, and where it needs guardrails.
Bedrock is where AWS runs models, retrieval and agents. AgentCore went generally available in October 2025, and Evaluations followed in March 2026. Bedrock Agents Classic closed to new customers on 30 July 2026. For engineering teams, the hard parts are agent identity, tool access and model retirements.
AI agentasks for data or an action
Scoped identityonly the access it needs
MCP serverthe one door into the platform
Logged, approveda person signs off on writes
Amazon Bedrockdoes the work
AgentCore Gateway is the MCP piece. It connects agents to MCP servers you already run, and it exposes APIs and Lambda functions as tools behind IAM or OAuth authorization. Each agent should get its own scoped credentials through AgentCore Identity. Log every tool call, and put a human approval in front of any tool that writes.
AgentCore Runtime, Memory and Observability
Agent code runs on AgentCore Runtime, with Memory for session context and Observability for traces. We wire the traces into your monitoring and set cost alerts. Both are done before users arrive.
Gateway and Identity
AgentCore Gateway turns APIs and Lambda functions into agent tools and connects to MCP servers you already have. AgentCore Identity lets agents act for a user, or as themselves, with scoped credentials. We design both with your security team.
Evaluations, Policy and Guardrails
AgentCore Evaluations, GA on 31 March 2026, and AgentCore Policy check quality and what an agent may do. Bedrock Guardrails cover content. We run evaluations in CI on every release.
Agents Classic and model lifecycle
Bedrock Agents Classic is closed to new customers, so new work goes on AgentCore. Models move through Active, Legacy and EOL states, so we keep a register and a test set per use case.
How an agent on Amazon Bedrock reaches your systems safely
Okta handles sign-in, and users then call an agent on AgentCore Runtime. It draws on Bedrock models and Knowledge Bases. Internal systems are reachable only through AgentCore Gateway, with AgentCore Identity issuing the credentials, while Guardrails and Evaluations watch what the agent says and does.
01 · Users
Oktaworkforce sign-in → Business app · OIDC sign-in
Business appportal · Slack · Teams → AgentCore Runtime · invoke agent
These are working documents, written against your configuration, so your team can run them without us.
document01
Use case brief with systems, data owners and success criteria
yours to keep
document02
Evaluation set from real past cases, run in CI
yours to keep
design03
Agent identity and tool access design reviewed by security
yours to keep
guide04
AgentCore deployment with Guardrails and Policy configuration
yours to keep
inventory05
Model lifecycle register with Legacy and EOL dates per use case
yours to keep
runbook06
Runbooks for on-call, model switches and access changes
yours to keep
talent and delivery · amazon bedrock
Certified engineers, wherever you need them.
Some clients want a project team, some want a couple of engineers on theirs. Either works. Onsite, nearshore or offshore, with handovers across time zones that keep things moving.
Forward deployed AI engineer (Amazon Bedrock, AgentCore)
AI platform engineer for AWS (IAM, VPC, AgentCore)
Retrieval engineer for Bedrock Knowledge Bases
LLM evaluation engineer
AI security engineer for agent identity
due diligence · amazon bedrock
Questions to ask any partner, including us.
Put these to anyone bidding on your Amazon Bedrock work. The ones who've done it answer differently from the ones who've read about it.
Q1
How will each tool call our agent makes be tied to an identity we can audit?
Agents acting with shared keys can't be reviewed or revoked cleanly. AgentCore Identity and Gateway exist for this, but someone has to design the roles.
Q2
What is in the evaluation set, and who wrote it?
An evaluation set built from real cases by the business owners is the only fair test. Synthetic prompts written by the vendor tend to flatter the agent.
Q3
What is the plan when our model moves to Legacy?
Bedrock retires models on a lifecycle it publishes. If you have no test set and no plan for switching, the retirement date becomes your outage date.
Q4
Should we migrate our Bedrock Agents Classic agents now or later?
Classic keeps running for existing customers. New investment is going into AgentCore, though, so it depends on how often each agent needs to change.
Q5
Which data will the agent see, and who approved it?
A knowledge base will index whatever you point it at. Give every document source an owner and access rules before it's indexed. Afterwards is too late.
01Should we build on Bedrock Agents Classic or AgentCore?
Bedrock Agents Classic closed to new customers on 30 July 2026. Existing customers can still use it, although AWS points to AgentCore for new work. So we build new agents on AgentCore and migrate the Classic ones as they come up for change.
02Which models can we use on Bedrock?
Several providers, including Anthropic Claude and Amazon Nova. The catalog changes too often for a fixed answer, so we read the current model cards and pick per use case, weighing quality and latency against cost and Region availability.
03What happens when a model we use is retired?
Models on Bedrock go from Active to Legacy to EOL, and there's a newer policy for models launched from 7 September 2026. We keep one evaluation set per use case. That way a replacement can be tested and switched before the EOL date.
04How long does it take to get an agent into production?
It depends on how many internal systems the agent calls, how access is approved and how much evaluation your risk team requires. The model work is rarely the slow part. Access reviews and test sets usually are.
05Do you embed engineers with our team?
Yes. Forward deployed engineers from Cohash sit with your business team, build on your real systems and data, and hand over evaluations, access controls and runbooks at the end. They can be onsite in the US, or nearshore and offshore in any time zone.
services
Related services
Projects on Amazon Bedrock usually end up touching one of these, so they're worth a look.
