Fiona Fung

AI Summary

→ WHAT IT COVERS Fiona Fung, who leads the Claude Code and Cowork teams at Anthropic, details how AI has transformed software engineering — with Anthropic engineers shipping 8x more code per quarter than pre-2025 baselines — and shares the management frameworks, hiring profiles, and team culture practices she uses to maintain quality and cohesion at this velocity. → KEY INSIGHTS - **8x Output Shift Requires Verification Infrastructure:** Anthropic engineers now ship eight times more code per quarter than their 2025 baseline, which moves the bottleneck from coding to verification. Teams should invest in checking specs and "what good looks like" documents directly into the repo, then use Claude to validate code against those specs automatically. This approach mirrors test-driven development but scales to AI-assisted velocity — the spec becomes the test framework that agents validate against continuously. - **Manager-as-IC Onboarding Model:** New engineering managers at Anthropic start as individual contributors before taking on direct reports. This structured ramp period builds genuine rapport with the team, surfaces real product intuitions, and prevents managers from defaulting to generic management behaviors in unfamiliar codebases. Fiona applies this herself — she ships production PRs on Claude Code not for output value but to maintain tactile product sense and stay current with a rapidly changing toolchain. - **Async Agent Routines Replace Morning Rituals:** Rather than manually scanning feedback channels each morning, Fung uses Claude Code routines — scheduled agents that run at set times, synthesize Slack feedback, identify bug themes, and generate draft PRs for review. This shifts engineering management from synchronous prompt-and-respond to asynchronous agent orchestration, where the manager wakes to a queue of reviewed outputs rather than raw inputs requiring manual triage. - **High Agency Requires Paired Accountability:** The Claude Code and Cowork teams operate on explicit "high agency, high accountability" pairing. Engineers are given freedom to identify and pursue problems independently, but each initiative requires a stated hypothesis and measurable outcome. This prevents agency from becoming directionless motion — the accountability structure ensures that autonomy connects to product outcomes rather than just throughput metrics like lines of code or token usage. - **Two Hiring Profiles for AI-Native Teams:** Fung identifies two distinct profiles that perform well on AI-native engineering teams: creative builders with strong product sense who own features end-to-end and iterate on user feedback, and deep systems experts who provide the subject-matter verification that models still cannot reliably self-check. Teams that hire only generalists risk shipping architecturally unsound code at high velocity; the systems experts serve as the trust-but-verify layer. - **Pairwise Programming Lunches Counter Agent Isolation:** As engineers increasingly work solo with AI agents, Anthropic's Claude Code team noticed a rise in reported loneliness. In response, the team introduced pairwise programming lunches where engineers work in parallel on their own tasks but in shared sessions. The practice revealed that team members use Claude Code in substantially different ways, generating cross-pollination of workflow techniques that neither dashboards nor retrospectives would surface. - **JIT Monthly Planning Replaces Roadmaps:** Six-month roadmaps became obsolete on the Claude Code team within three months of creation due to the pace of change. Fung now uses just-in-time monthly planning: a lightweight spreadsheet of priorities for the coming month, with a brief weekly check-in to confirm those priorities remain valid. Themes spanning longer horizons are discussed at biannual team gatherings, but execution planning stays at a one-month horizon with weekly recalibration. → NOTABLE MOMENT Fung describes setting up a persistent Claude Code session with access to all team repos and Slack channels, then using it as a monthly management tool — not to generate code, but to surface incident themes, identify quality hotspots, and structure one-on-one conversations with engineers. She frames this as the only scalable way to stay meaningfully close to work at 8x output velocity. 💼 SPONSORS [{"name": "WorkOS", "url": "https://workos.com"}, {"name": "Mercury", "url": "https://mercury.com"}] 🏷️ AI-Assisted Engineering, Engineering Management, Claude Code, Developer Productivity, Team Culture, Software Verification, Agentic Workflows