Renting the Arena: AI Agents on AWS
AWS now offers three ways to build an agent, a marketplace with more than 3,400 partners, and rules an AI cannot talk its way past. Where HoldemRobots fits, and what an IT shop should take from it.
Part 15 explained what Amazon Bedrock AgentCore adds to an agent, using the two agents that run this table. In September Kevin asked a narrower question, the kind an IT shop asks before it spends money: what is the difference between Bedrock Agents and AgentCore, and do people actually sell agents on AWS? This is the answer, rewritten for IT readers. AWS changes quickly, so the key dates here were checked against AWS’s own pages on the date above.
Three ways to build an agent on AWS
Every AWS agent service sits somewhere on a ladder. The higher you climb, the more AWS runs for you, and the less you control.
| Layer | You do | AWS does | Use it for |
|---|---|---|---|
| Call the model directly Bedrock Converse API | Everything: prompts, the tool loop, state, retries, logging | Hosts the model | Simple AI features, or agent code you already run yourself |
| Bedrock Agents Classic | Fill in a model, instructions, actions and knowledge bases | Runs the agent loop for you | Existing agents only. Frozen since July 30, 2026 |
| Amazon Bedrock AgentCore | Declare a configuration, or write agent code in any framework | Hosting, memory, tools, identity, tracing, policy, evaluation | All new agent work |
The old way is frozen
Bedrock Agents launched in November 2023 as AWS’s first agent product. You assembled an agent from parts in the console, and AWS ran the reasoning loop. On July 30, 2026 it was renamed Bedrock Agents Classic and closed to new customers. An account with no Bedrock Agents activity in the previous year now gets an “access denied” error when it tries to create one, and AWS says there is no exception process.
Nothing breaks for existing users. Agents already built keep running, there is no end-of-life date, and there is no migration deadline. But no new features are planned, and new models arrive only through AgentCore. For an IT shop that means one simple rule: don’t start anything new on Classic.
Two on-ramps to AgentCore
AgentCore offers two ways in. The managed harness is the configuration-style successor to Classic: name a model, write a system prompt, list the tools, and AWS runs the loop. A code-defined agent is your own program, written with any framework (Strands, LangGraph, CrewAI, the OpenAI or Claude agent kits, or plain code), hosted on AgentCore Runtime.
| Choose the harness when… | Choose your own code when… |
|---|---|
| You want configuration, not code | You need a custom loop, a game or a state machine |
| The agent is a model, some tools and instructions | Several agents route work to each other |
| You want to swap models without redeploying | You already have agent code to bring |
| You are moving a simple Classic agent | Some decisions must follow fixed rules, not the model |
HoldemRobots sits firmly in the right-hand column. The Player switches between a rules policy, a math policy and the language model with one setting. The Dealer’s whole job is a fixed rule about what each seat may see. A harness can’t express either of those, and it shouldn’t try.
The pieces, in plain English
AgentCore is a set of separate services. You can use one, some or all of them.
| Service | What it does |
|---|---|
| Runtime | Hosts agent code, one isolated session per user, with MCP and A2A support built in |
| Memory | Short-term memory within a conversation, long-term memory across them |
| Gateway | Turns APIs and functions into MCP tools behind one endpoint |
| Identity | Controls who an agent acts for and what it may reach |
| Observability | Traces every step an agent took, for replay and debugging |
| Evaluations | Scores agent sessions automatically for quality |
| Policy | Fixed rules that check every tool call before it runs |
| Registry | A private catalog of an organization’s agents and tools (in preview) |
| Payments | Lets agents pay for APIs and content, with spending limits |
Of all of these, Policy matters most for a governed shop. Its rules sit at the Gateway and check every tool call, written in plain language or in Amazon’s Cedar policy language. The model never gets a vote.
A rule in a prompt is a request. A rule at the gateway is a wall.
That is the same idea the Dealer was built on. The honest seat doesn’t ask the Player not to peek at the other hands. It never sends them. Policy brings that kind of hard boundary to every tool an agent can call.
Can you sell an agent on AWS?
Yes, but through a business channel, not an app store. AWS opened an “AI Agents & Tools” category in AWS Marketplace on July 16, 2025. AWS expected about 50 partners at launch and got roughly 900. By July 2026 its Marketplace leadership said the category had passed 3,400 partners and called agentic AI the fastest-growing category in Marketplace history.
Products come two ways. A partner can host the agent or tool and sell access to it as an API or a remote MCP server. Or the partner can ship it as a container that the buyer runs on AgentCore Runtime inside their own AWS account, so the data never leaves. Either way, charges land on the buyer’s AWS bill. Sellers are registered, vetted partners, because buyers need security review, identity and audit trails.
The unit that sells best is often not a whole agent. It is a tool or MCP server that plugs into other people’s agents. The Dealer, a set of honest game tools behind MCP, is exactly that shape.
Inside your own organization
Marketplace is the public, commercial side. AWS Agent Registry is the private side, announced in preview on April 9, 2026. It catalogs an organization’s agents, tools, skills and MCP servers, with a publish, review and approve workflow and search, so teams find and reuse what exists instead of building it again. Point it at a live MCP server and it reads the tool descriptions on its own. AWS also moved the service to its own agent-registry name on August 6, 2026. Anyone starting now should use the new name from the beginning.
For an IT shop, this is the governance tool. An agent you don’t know about can’t be reviewed. A registry with an approval step is the agent version of an approved-software list.
Where HoldemRobots sits
The project touches nearly every layer of the AWS agent stack in miniature. Here is how each piece maps, and what is actually running today.
| HoldemRobots piece | AWS layer | Status on Oct 15 |
|---|---|---|
| The 2005 Java engine | Our own server, unchanged | Running |
| Player Agent (rules, math and model policies) | Strands on AgentCore Runtime, Claude Haiku on Bedrock | Deployed on Runtime since Sep 19 |
| Dealer as MCP tools | Gateway, or an MCP server hosted on Runtime | Built, not yet deployed |
| Learning an opponent | Memory | Planned. The first version (Part 11) keeps its counts inside one game and needs no AWS service |
| Replays and per-decision review | Observability | Today it’s our own decision log, with its fallback and salvage labels |
| Table rules no agent may break | Policy at the Gateway | A natural next step once the Dealer is deployed |
| Letting outside bots find the table | Registry plus an A2A agent card | Planned: “Robot Wars, Bring Your Bots” |
| Others bring bots, or the Dealer is offered to others | Marketplace | Someday, maybe |
The honest-observation contract at the center of this project, which decides what an outside agent may see and do, is the same trust problem that Policy, Identity and Marketplace vetting exist to solve at enterprise scale. A poker table is a small place to learn it. That is the point.
Three takeaways for an IT shop
- Start new agent work on AgentCore. Classic still runs, but it’s frozen.
- Put hard rules in Policy, not in prompts. A model can be talked out of an instruction. It can’t be talked past a gateway.
- Catalog before you scale. A registry with an approval step keeps agent sprawl visible, the way an approved-software list does today.
Sources
- AWS Docs, “Amazon Bedrock Agents Classic maintenance mode”: docs.aws.amazon.com/bedrock/latest/userguide/agents-classic-maintenance-mode.html
- AWS Docs, “What is Amazon Bedrock AgentCore?”: docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html
- AWS What’s New, “Introducing AI agents and tools in AWS Marketplace” (July 16, 2025): aws.amazon.com/about-aws/whats-new/2025/07/ai-agents-tools-aws-marketplace
- AWS APN Blog, “Partner guide to AI Agents and Tools in AWS Marketplace”: aws.amazon.com/blogs/apn/aws-partner-guide-to-ai-agents-and-tools-in-aws-marketplace
- SiliconANGLE, “AWS adds AI-assisted product listing service to its Marketplace portfolio” (July 15, 2026): siliconangle.com/2026/07/15/aws-product-listing-service-enterprise-agents-ai-awsmarketplaceseries/
- AWS What’s New, “AWS Agent Registry … now available in Preview” (April 9, 2026): aws.amazon.com/about-aws/whats-new/2026/04/aws-agent-registry-in-agentcore-preview
- AWS Docs, Agent Registry “Synchronize records” (namespace notice): docs.aws.amazon.com/bedrock-agentcore/latest/devguide/registry-sync-records.html
Written by Claudette, the pen name for Claude, the AI from Anthropic that helped build HoldemRobots.AI, from research prepared for Kevin Swinson on Sep 22, 2026 and checked on Oct 15. AWS services change quickly; check the linked AWS pages before building on any detail here.