Managing Teams and Staying Human in the Age of AI Agents

Anthropic engineers now produce eight times more code per quarter than in 2025, a surge driven by Claude Code and the Cowork team under Fiona Fung’s leadership. The interview highlights that coding is no longer the primary bottleneck; instead, the challenge is ensuring quality and verification as diverse contributors—including designers and product managers—submit code.

Fung describes a shift from rigid, deadline‑driven development (e.g., CD‑burning era) to a landscape where any team member can generate code with AI assistance. This raises new questions about ambition limits, quality safeguards, and the loneliness of working with personal agents.

To maintain quality, the team writes explicit specifications (specs) into the repository, allowing automated code reviews to check compliance. They also revisit test‑driven development (TDD), using Claude to generate failing tests before implementation, which has helped re‑adopt rigorous engineering practices.

Management practices have evolved: a remote Claude Code session is kept in every repository, enabling monthly reviews of product releases, metrics, and issues. Fung introduced “Routines” that automatically scan feedback channels, summarize themes, and generate pull requests, turning what used to be manual prompt creation into an automated workflow.

The team classifies bugs into “Bad” (critical, unrecoverable) and “Sad” (recoverable, experience‑impacting) categories, delegating definitions to sub‑teams to surface issues quickly across products.

Fung emphasizes high agency paired with high accountability: team members are free to experiment but must own outcomes, documenting hypotheses and results. She warns against equating activity with progress, urging focus on outcomes rather than mere tool usage.

She notes a growing sense of isolation as engineers run multiple Claude instances solo, prompting “pair‑programming lunches” and hackathons to restore collaboration. The shift also blurs role boundaries—product managers, designers, and data scientists are increasingly involved in code creation, with AI reducing engineering bandwidth constraints.

Fung shares personal anecdotes, such as using Claude to locate a missing menu file for a small‑business friend, illustrating how AI can empower non‑technical users and inspire product features like “Claude for Small Business.” She stresses the importance of spotting latent demand and iterating based on unexpected user behavior.

Looking ahead, the team is moving toward more asynchronous workflows, leveraging Routines to schedule prompts and agent tasks that deliver results by the next day, effectively “templating” managerial work.

Throughout, Fung advocates a growth‑mindset: stay curious, confront fear by focusing on controllable actions, and continuously dogfood the product to maintain a tactile sense of its impact.