Shipping Products When You Don't Know What they Can Do — Ben Stein, Teammates

Introduction and Teammates Overview 00:14

  • The speaker is a founder of Teammates and will focus on a product management perspective.
  • Teammates is a platform designed for creating and managing digital workforces, focusing on collaboration between humans and computers.
  • Users can design personalized AI teammates with unique avatars and personalities that integrate into digital tools like Google Workspace and Slack.

The Challenge of Product Unknowns 01:59

  • A customer’s request to tag an AI teammate in a Google Doc comment highlighted uncertainty about the product's actual behavior.
  • The speaker acknowledges not always knowing what the product can do, which inspired the talk.
  • This uncertainty is due both to the unpredictable nature of large language models (LLMs) and the limitless ways customers might use a flexible, agent-based product.

Evolving Product Management for LLM-based Products 03:07

  • Product management is undergoing a major shift due to the opacity and emergent behavior of AI-based systems.
  • Traditional product development assumed a well-understood foundation and clearly defined user boundaries—now both are unclear.
  • Products built atop LLMs have unknown capabilities, and open-ended user interfaces (like free textboxes) invite unpredictable uses.

Rethinking Feature Design: Affordances, Not Requirements 06:04

  • Shift focus from specifying exact requirements to outlining affordances—what the agent is allowed or able to do.
  • Product managers need to define building blocks and enable emergent behaviors rather than trying to predict every possible outcome.
  • Behaviors emerge unpredictably, so discovering functionality becomes an ongoing process.

Emergent Functionality and Communication Challenges 07:14

  • The job of product managers grows to include identifying new, unexpected capabilities as they arise in use.
  • It’s hard to communicate or specify emergent behaviors with traditional tools like PRDs or Figma.

Evals as the New Specification 08:08

  • Evals (evaluation frameworks) are used to test and measure the probabilistic outcomes of AI agents, such as whether an AI responds with the right tone or style.
  • Evals become a living specification for what the product does and can serve as an ongoing measure of agent performance.
  • Product people should engage with evals directly for better understanding and documentation of product behaviors.

The Role of “Vibe Coding” and User Feel 10:46

  • Prototyping and “vibe coding” (experimenting quickly to see how an agent feels in real use) are critical because the right experience is hard to predict on paper.
  • User feedback often highlights unexpected annoyances or delights that specs would miss, reinforcing the need for fast iteration.

Shifting QA and Bug Reporting Approaches 13:02

  • Testing AI agents is about discovering what they actually do, not just verifying pre-defined requirements.
  • Traditional bug tracking struggles to classify issues in probabilistic systems—distinctions between features and bugs blur.
  • Acceptable performance can be defined in probabilistic terms (e.g., 90% success on key behaviors), and thresholds in evals serve as new criteria for shipping.

Redefining Customer Communication and Collaboration 16:04

  • Traditional product management roles—visionary and honest broker—are harder to play when both current functionality and future direction are uncertain or seem unbelievable.
  • The most effective strategy is to position the customer relationship as co-inventing the future, setting expectations for shared discovery and experimentation.
  • Customers not ready for this collaborative, evolving process may not be the right fit at this stage.

The Future of Product Management with AI Agents 18:13

  • The speaker expresses excitement and surprise at the rate of emergent behaviors and improvements as models improve.
  • Product management and development disciplines will need to rapidly adapt and let go of many legacy methodologies.
  • Core product principles remain (e.g., listening to customers), but techniques and processes are being transformed by LLM-based and agent-driven products.