Craig Wattrus is a South African product designer, technologist, and member of technical staff at Obvious, developing interfaces that help AI agents collaborate effectively with people. His work on Transform Agent and agent-driven data onboarding addresses a practical design problem: how autonomous software can take useful initiative while keeping human decisions visible, reviewable, and reversible.
From payments to AI product design
Wattrus studied computer science at the University of Johannesburg and began his career at Thoughtworks. He subsequently worked at Yoco on tools for small businesses and at Braintree on developer experience, payment integrations, and documentation for Braintree and PayPal. His work improving Braintree’s support experience emphasized contextual assistance, measurable outcomes, and access to human help.
He later worked on products at Tegus, Foxtrot, and Crafty before joining Flatfile, now Obvious. As product lead for Transform Agent, he helped develop context-aware data preparation that incorporates schemas, validation rules, and previous transformation patterns; users can inspect suggested repairs, approve changes, and reuse transformations.
Designing useful AI coworkers
Four form factors for AI coworkers. Wattrus distinguishes invisible agents operating behind the scenes, ambient systems surfacing opportunities, inline assistance embedded in existing workflows, and conversational tools for open-ended collaboration. His AI Engineer World’s Fair talk illustrates how each interaction model requires different levels of visibility and human oversight.
Agent character as product behavior. Using Vercel’s v0, he built a prompt-tuning interface to test how an orchestrator’s system prompt affected friendliness, caution, initiative, and communication. His goal is an agent that advances the work without exceeding its mandate or burying users in unnecessary explanation.
Legible, reversible autonomy. His prototypes explore collaborative AI cursors, agents operating across multiple files, visual explanations, explicit approvals, snapshots, and rollback. One experiment combined related JSON and CSV files and identified duplicates; another recognized that missing employee information required human intervention.
Functional prototypes over static mockups. Drawing on his woodworking practice, Wattrus treats models as materials whose capabilities and limitations become apparent through direct experimentation with working software.
His independent projects include Inby, a Mac-native workspace combining terminals, agents, browsing, and writing; Bochi, a stationery-focused catalog; and Maats, an experimental studio spanning software, handcrafted objects, and sustainable materials.