To make a high-agency culture successful, it is crucial to balance freedom with responsibility. At Anthropic, teams have the autonomy to pursue their ideas, but they are also held accountable for the outcomes and the hypotheses they are testing, ensuring freedom is directed toward impactful work.
The next frontier for engineering work is asynchronous management. Instead of synchronous prompting, managers create "routines" that automatically run daily, analyze feedback channels, identify issues, and even generate draft PRs for review. This moves management to a higher level of abstraction.
To ensure managers deeply understand the current tooling, codebase, and team dynamics in a rapidly changing AI environment, they are required to onboard as ICs. This player-coach model builds rapport and grounds their leadership in direct, hands-on experience before they begin managing people.
Engineers struggling with the shift to AI are often driven by fear of obsolescence. The solution is to encourage a growth mindset, lean into the fear, and identify concrete actions within their control. This shifts the narrative from "happening to me" to "happening for me," turning frustration into agency.
An unexpected side effect of engineers working primarily with AI agents is loneliness. To foster team connection and shared learning, Anthropic started pairwise programming lunches. This helps teammates see each other's unique AI workflows and restores a sense of community.
Major product opportunities are revealed by observing how customers use your product in unintended ways or "jump through hoops" to achieve a goal. For example, Anthropic noticed non-engineers struggling to use their coding tool, revealing the latent demand for CoWork, a knowledge-work assistant.
To scale code review with 8x output, teams should codify and check-in their standards—specs, design systems, style guides—into the repository. AI reviewers can then automatically validate new code against this explicit "statement of what good looks like," reducing the burden on human reviewers.
AI tools have removed coding as the primary bottleneck in software development. The new ceiling for an engineer's impact is their ambition and ability to conceptualize big ideas, as execution has become drastically easier. It's no longer about what can be built, but how big you can think.
To manage the 8x increase in code shipment, managers use AI agents with full repo and communication access. This AI summarizes shipped products, feedback, and metrics, enabling data-driven conversations about impact, learnings, and areas for investment, replacing a previously manual process.
Traditional roadmapping is too slow for the pace of AI development. Anthropic's team uses a "Just-in-Time" planning model: a simple spreadsheet outlining priorities for the next month, with a quick check-in each week to ensure it's still relevant. This prioritizes adaptability over long-term prediction.
As AI handles routine coding, the most valuable engineers are either "dreamers" with strong product sense who can own features end-to-end, or deep subject matter experts who can verify and handle the complex, trust-critical parts of the system where human verification is still essential.
To maintain a quality bar across diverse products, use a simple framework. "Bad" errors are critical and irrecoverable (e.g., a crash), while "Sad" errors are recoverable annoyances (e.g., UI flicker). Each team defines what constitutes Bad vs. Sad for their area, enabling a high-level, comparable view of product health.