So an AI walks into your poker game... hear us out

Summary

The episode covers three applications of generative AI: research into scorpion antivenom, decision models for fast and bounded choices, and governance platforms for managing AI agents across an organization.

1. AI-assisted research into scorpion toxins

The hosts discuss an Anthropic blog post about using an AI model to research scorpion neurotoxins. They contrast this approach with traditional antivenom production, which involves collecting venom from many scorpions and producing antibodies in horses.

According to the hosts’ description of the blog post, the model worked largely on its own for about 64 hours. It searched protein databases, selected targets, considered roughly 300,000 design options, and narrowed them to candidates. Seventeen candidates reportedly bound to toxins. The hosts stress that this is a research milestone, not a finished treatment.

2. Decision models: choosing among options rather than generating prose

Guest Mike Reinhold explains how decision models differ from conventional large language models (LLMs):

• Conventional LLMs generate text token by token, often producing an explanation or open-ended response.
• Decision models use an LLM-based model with a decision-oriented output layer. Instead of generating prose, they return probabilities across a defined set of choices.
• Because the choices are bounded and can be evaluated in one pass, decision models can be faster and less expensive for suitable tasks. In the example discussed, Strands Decider took about 147 milliseconds, compared with about 3.5 seconds for a conventional LLM response.
• Their probabilities come from the model’s decision output, rather than from asking a text-generating LLM to write a number that looks like a probability. These probabilities can indicate confidence, but their usefulness still depends on the quality and completeness of the input data.

The poker demonstration illustrates the concept: a model can choose among actions such as calling or folding rather than explain its thinking in paragraphs. Other examples include selecting a direction in Snake, making yes-or-no assessments, evaluating an agent’s response, or deciding whether an agent should proceed with an action.

Decision models may be useful when:

• The possible outputs are limited and clearly specified.
• A fast route to an appropriate model or workflow is needed.
• Another model’s or agent’s response needs to be scored or evaluated.
• Confidence-based controls are needed—for example, automatically allowing high-confidence actions while sending uncertain or risky ones for human review.
• A proposed tool call or action needs to be checked for authorization.

Mike characterizes conventional LLMs as useful for deeper reasoning, planning, and tool orchestration, while decision models are useful for quick, calibrated choices. The two can be combined: an LLM can handle open-ended reasoning, while a decision model helps assess, route, or gate actions.

The Strands Decider models are described as open source and available through GitHub and Hugging Face. They can run locally on a CPU or GPU; cloud GPUs can provide more performance and scaling for larger workloads.

3. Managing agentic AI across an organization

Guests Jen and Melanie demonstrate an AI management platform intended to help organizations govern, monitor, and scale agents. They frame the need around common problems: security teams may not know which tools are in use, developers may lack approved patterns, and business leaders may be surprised by costs.

The platform’s capabilities, as shown in the demonstration, include:

• Organization-wide visibility: A dashboard provides a consolidated view of applications, users, departments, projects, environments, budgets, and policy violations.
• Governance of models and tools: Organizations can define approved models, tools, knowledge bases, and policies. The guests say this can prevent unapproved or vulnerable tools from being used across projects.
• Cost controls: Budgets can be set at the project level and further limited by agent, environment, or model. Changes go through review and are recorded, while remaining adjustable.
• Reusable blueprints and harnesses: Developers can start from organization-approved templates that specify elements such as tools, knowledge, memory, architecture, and assessment criteria, rather than building every agent setup from scratch.
• Quality checks and approvals: Projects are assessed against defined criteria. Failed checks prevent promotion to production; successful results provide evidence for an approver to review before wider deployment.
• Integration with managed agents: The demonstration also shows a connection to Bedrock managed agents, described as being in preview. The example uses a managed agent with a selectable runtime and configurable model, skills, and MCP tools. A sample incident-investigation agent retrieves CloudWatch information and produces a report.

The broader message is that agent development needs both flexibility for builders and a control plane for security, quality, cost, and operational oversight.

Speakers and sources featured

• Jay — host.
• Bharathi — co-host; the auto-generated subtitles render her name inconsistently as “Barty” and similar variants.
• Mike Reinhold — AWS enterprise architect and guest discussing decision models.
• Jen — guest presenting the agent-management platform; her name is rendered inconsistently in the subtitles.
• Melanie — guest presenting the platform and managed-agent demonstration.
• Vivek Raja — AWS Hero referenced by the hosts for work on model routing; not a featured speaker in the episode.
• Anthropic blog post — cited in the discussion of AI-assisted scorpion-toxin research.