Skip to content

Alfred PoindexterPrincipal product designer

I design and build AI agent products people can supervise, evaluate, and trust.

5%sitewide conversion lift

measuredPowerWizard. Homepage hero test, driven by mobile.

Behavior, psychology, design, technology

I design for the person in the loop.

Managing agents turns out to look a lot like managing people: you define roles, set boundaries, check the work, and decide what stays on your own desk.

From Essay 01 in The Machine Org Chart

A print of a human head in profile, built from teal, navy, coral and marigold squares that scatter toward the back of the head, inside navy contour lines on cream paper.

Selected work

Case studies with interactive walkthroughs

Two agent products I designed, built, and evaluated, and a consumer energy marketplace measured in conversion tests. Each one is shown through its own interface.

Project Aegis

I designed, built, and evaluated a browser-agent shopping audit with a separate deterministic verifier and a read-only audit console.

Read the case study

PowerWizard

Consumer energy marketplace work spanning usability research, conversion tests, and reducing design rework.

Read the case study

How I work, in four chapters

A field of decisions

Each square is a decision someone makes. Over four chapters they sort from behavior to oversight, in the order I design.

  1. 01

    Behavior

    I start with what people do. For PowerWizard that meant a 20-person usability study in Texas and a 15-person rooftop-solar study in California.

    PowerWizard

  2. 02

    Mental models

    Then I design for the model a person forms of the system. At every checkpoint, advice never looks like approval, and progress shows only from real recorded events.

    Essay 01, checkpoint patterns

  3. 03

    Orchestration

    “Divide authority, not just labor.” In Agency OS, a strategist agent, a producer agent and a reviewer agent each draft their own part of the work.

    Agency OS

  4. 04

    Oversight

    Every draft passes one person, who approves the composition. In Project Aegis the shopper agent stops before checkout, payment, and account creation, and a separate verifier grades the run.

    Agency OS, Project Aegis

Practice

Human-centred, then machine-shaped

Six human-centred steps are the roots of how I work. Today they feed one loop: I research, design, prototype in code, and evaluate what I build.

ToolsFigmaCursor & coding agentsLangSmithRive

Roots: the human-centred process

  1. 01ContextualizeUse analytics to understand how people engage with the product.
  2. 02EmpathizeObserve behavior, find the cognitive friction, gather support and engineering data.
  3. 03DefineSynthesize research into a point of view on root causes and unmet needs.
  4. 04IdeateGenerate options with engineers, designers, business leaders, and end users.
  5. 05PrototypeBuild something a person can evaluate in context.
  6. 06TestCheck whether the solution meets the user’s needs, and the business’s.

Today’s loop

  1. 01Research
  2. 02Design
  3. 03Prototype in code
  4. 04Evaluate

Writing

I publish The Machine Org Chart

Human intent isn’t friction to engineer away. It’s the spec.

From Essay 01, Why Generalist AI Agents Fail in Production

The Machine Org Chart

Notes on building and managing AI agent teams that hold up in production.

Essay 018 min read

Why Generalist AI Agents Fail in Production (And What to Build Instead)

The future of production-grade AI isn’t one all-powerful agent. It’s a structured digital workforce: clear division of labor, governed by human intent.

Read the essay

Three failure modes

  • They grade their own homework
  • Their permissions grow faster than anyone’s understanding of them
  • They bury the human decision
Human:intentandapprovalsdelegationstopsCoordinatorStrategistProducerReviewerOwntools·ownbudget