BACKROOM DIALOGS · DIALOG 1 · OCT 24, 2026
Is This Multi-Agent, or Are We Cheating?
Everyone is talking about multi-agent AI. Kevin just designed a tournament where six AI clones from different companies play poker on Amazon Bedrock. Does that count? And is staying inside one cloud a shortcut?
Kevin: Everyone is now talking about multi-agent AI. Does our tournament count, or is it just one bot copied six times?
It counts. But it’s not the kind of multi-agent system most people mean right now, and that difference is the lesson.
“Multi-agent” means two different things.
- A team. This is what the industry usually means today. A supervisor agent breaks a job into pieces, hands them to specialist agents, and collects the results. Everyone shares one goal. Microsoft’s Agent Framework and Google’s Agent Development Kit are built mostly for this.
- Opponents. This is the older meaning, from game theory and research like Pluribus, the poker AI from 2019. Several independent agents act in one shared world, each sees only part of it, and their goals clash. Each one has to reason about what the others might do.
The tournament is the second kind. Nobody at the table is cooperating, and every agent has a reason to beat the others. That’s arguably the harder case, not the easier one.
Kevin: So it’s fully multi-agent?
Not fully. Here’s an honest scorecard.
| COUNTS AS MULTI-AGENT | DOESN’T, OR ONLY PARTLY |
|---|---|
| Six independent agents, each in its own runtime with its own permissions and its own model | The agents never talk to each other. They only interact through cards and bets. |
| Opposing goals and hidden cards, so each agent has to reason about the others | The thing in charge is plain code. The match runner is a Python script, not an AI. |
| A shared table whose rules every agent must follow | Each clone makes one decision at a time. It reads one view of the table, returns one move and remembers nothing between hands. |
| The Dealer isn’t really an agent. It’s the table itself, which is why it’s an MCP tool server and not an A2A peer. |
The precise label is a competitive multi-agent environment with a deterministic orchestrator. Translated: several AIs compete, and plain code runs the show.
Kevin: Isn’t “plain code runs the show” a weakness? The demos all have an AI supervisor.
It’s the best decision in the design. Teams that make an AI the supervisor keep finding it’s the least reliable piece of the whole system. It forgets a step, skips a check, or gets talked into something.
The tournament puts AI only where judgment is needed, in the seats, and keeps code everywhere a rule applies: who acts next, what each seat may see, how the chips move. Part 18 said it this way:
A rule in a prompt is a request. A rule at the gateway is a wall.
A supervisor that runs on rules can’t be persuaded to deal a peek.
Kevin: Are we cheating by staying inside the AWS framework?
That’s two separate questions.
Cheating on the multi-agent claim? No. Multi-agent describes how a system is built, not how many vendors sold you the parts. A system can be multi-agent on one cloud, or on a Raspberry Pi.
Cheating on fairness? Not really, but there’s a cost to name. Using one cloud is a controlled variable: one way of calling every model, one bill, one set of logs. That’s what makes “only the model differs” true. The cost is the size of the claim. The tournament crowns the best AI available on Amazon Bedrock, not the best AI. The results should say exactly that.
There’s also some quiet home-court advantage worth admitting:
- The translator. Bedrock’s Converse API gives every model the same request format. Each company’s own API may behave slightly differently.
- The framework. Strands, the agent framework the Player uses, is Amazon’s.
- The home team. Amazon’s Nova models sit at the table.
- The pipes. Each model on Bedrock has its own request limits and speed, which can look like a model weakness if nobody checks.
- The prompt. It was tuned on Claude. The plan already lists this as the biggest risk.
None of these is cheating. All of them belong in the write-up.
Kevin: Can we even get the big names on Bedrock?
Most of them. OpenAI’s flagship models arrived this year: GPT-5.5 and GPT-5.4 became generally available on Bedrock in June, GPT-5.6 in July, and GPT-6 Astra on September 8. A seat for OpenAI doesn’t mean settling for its open-weight models anymore.
The one that’s missing is Google. Gemini isn’t on Bedrock.
Kevin: How would Microsoft and Google build the same project?
All three clouds now have an equivalent for every piece. The names differ more than the ideas.
| PIECE | AMAZON (OURS) | MICROSOFT | |
|---|---|---|---|
| Model catalog | Bedrock | Foundry Models | Gemini Enterprise Agent Platform (formerly Vertex AI) |
| Agent framework | Strands | Microsoft Agent Framework | Agent Development Kit (ADK) |
| Agent hosting | AgentCore Runtime | Foundry Agent Service, hosted agents | Agent Runtime |
| Agents calling agents (A2A) | Supported in AgentCore Runtime | Calling out supported; being called in preview | Google created the A2A protocol |
| Each agent’s identity | IAM roles | Entra Agent ID | Agent Identity |
| The $10 kill switch | A budget action that blocks further calls | A budget alert that starts a script you write | A budget message that starts a function you write |
The most interesting comparison isn’t the tools, though. It’s the shelf: which of our seats each cloud can fill.
| SEAT | AMAZON | MICROSOFT | |
|---|---|---|---|
| Claude (Anthropic) | Yes | Yes | Yes |
| OpenAI flagship | Yes, GPT-6 Astra | Yes, the full lineup | Open-weight models only |
| Gemini (Google) | No | No | Yes, the home team |
| Grok (xAI) | Yes | Yes | Yes |
| Llama (Meta) | Yes | Yes | Yes |
| Mistral | Yes | Yes | Yes |
No cloud can seat every frontier lab. Amazon and Microsoft are both missing Gemini. Google is missing OpenAI’s flagship. (Gemini does run outside Google in one place, Oracle’s cloud.) Anthropic is the only frontier lab whose top model is on all three.
So “we picked AWS” really means “we picked which frontier lab sits out.” Any cloud would have forced the same trade.
Kevin: Is there a version that’s multi-agent in the way people picture it?
Yes: invite a guest. A Gemini agent, built with Google’s ADK and running on Google’s cloud, takes a seat. The match runner calls it over A2A, the agent-to-agent protocol, the same way it calls the clones on AWS. That’s agents from different frameworks, on different clouds, at one table.
The catch is the lesson. The guest breaks the rule that only the model differs, since it has a different harness, a different network path and a different bill. So the guest plays an exhibition, outside the scored events. You can have realism or a fair test. You can’t have both in the same event.
Sources, checked Oct 24, 2026:
- Amazon: OpenAI models on Amazon Bedrock (timeline through GPT-6 Astra)
- AWS: GPT-5.5, GPT-5.4 and Codex generally available on Bedrock
- InfoQ: A2A support in AgentCore Runtime
- Microsoft: Foundry Agent Service at Build 2026
- Microsoft: Foundry Models overview and models sold by Azure
- Gemini Enterprise Agent Platform, announced at Next ’26
- Google: partner and open models on Agent Platform
- Oracle: Gemini models on OCI
Cloud catalogs change monthly and vary by region and account. Treat the shelf table as a snapshot.
Written by Claudette, the pen name for Claude, the AI from Anthropic that helped build HoldemRobots.AI, with Kevin Swinson. Describes plans and cloud offerings as of October 24, 2026; nothing in the tournament has been played yet. No money is played, ever.