Tech • AI • Robotics
How the Claude Code team uses Claude Code
TL;DR
Claude Code users inside Anthropic increasingly delegate most software work to cloud-based agents, shifting from line-by-line supervision to higher-level goals, while rapid model progress forces constant product redesign.
KEY POINTS
From coding assistant to autonomous agent
Workflows that once centered on prompting a model to write a class or function have expanded into broader tasks such as verification, code review, brainstorming and monitoring. Team members now describe giving the system an outcome to achieve rather than supervising each tool call, with one estimate placing 70% to 80% of daily work inside Claude Tag, the Slack-native interface layered on top of Claude Code.
Context access improves decisions
A major advantage of the Slack-based setup is access to organizational context, including product decisions, past discussions and team priorities. That wider context is used to improve judgment on more complex assignments, allowing the system to handle work that goes beyond isolated implementation tasks and into product-aware decision-making.
AI product cycles are shrinking sharply
Engineers described a major shift in software development economics: technologies that once had a shelf life of years may now be overtaken in months. Building on frontier AI requires balancing immediate customer value with preparation for capabilities that are advancing every two months or less, making roadmap planning more volatile than in traditional software products.
Useful features can become obsolete fast
One example was the early reliance on to-do lists to help models complete long-horizon work. Around the era of Sonnet 3.5, those lists significantly improved performance because models often dropped tasks before finishing them. As memory and coherence improved, that aid became far less necessary, illustrating why teams must stay detached from features designed mainly to patch temporary model weaknesses.
Interactive tools gave way to richer interfaces
An ask-user-question tool was initially important for making the agent interactive, but later improvements reduced dependence on it. In many cases, richer outputs such as HTML artifacts with diagrams and mockups now gather feedback more effectively, showing how interface layers are moving beyond raw chat and visible transcripts.
Cloud execution enabled persistent loops
Running agents locally created a simple problem: work stopped when a laptop closed. That pushed development toward remote boxes, then hosted containers and web-based environments where agents can continue operating in the background. Persistent cloud execution opened the door to recurring routines or “loops,” such as reviewing daily feedback, grouping issues by importance and fixing high-confidence problems automatically.
Code review is shifting from nits to architecture
Automated review now handles many low-value comments that once signaled a human had read the pull request. The more valuable human role is increasingly focused on larger questions such as service boundaries, API structure and product intent. In practice, AI review is being used to direct attention toward the changes most likely to matter rather than replacing human oversight altogether.
Fan-out workflows add confidence through compute
One of the most important internal patterns is a fan-out workflow that explores many possibilities, then filters results back for human use. In code review, that can mean searching broadly for bugs and then running adversarial checks from multiple perspectives before surfacing only the strongest findings. The same pattern can be applied to performance analysis, research and planning, effectively using heavy inference-time compute to raise confidence.
Deterministic orchestration builds trust
A notable design choice is letting the model write code that coordinates sub-agents. That creates a hybrid of deterministic software behavior and probabilistic model behavior, which users say improves trust because loops, iterations and routing logic are explicit and repeatable even when the underlying reasoning remains agentic.
Claude Tag is being used to build itself
Internal teams are aggressively using Claude Tag to develop Claude Tag, including setup of development environments, implementation, testing and integrated verification. In one example, a new internal tool moved from idea to stakeholder discovery, mockups, implementation, deployment, event tracking and feedback monitoring largely through Slack, with the system also tagging the developer when new user feedback arrived.
The engineer’s role is moving up a level
The biggest cultural shift is from manual craftsmanship toward idea generation, evaluation and orchestration. Tasks once associated with individual mastery, from performance tuning to pixel-perfect interface work, are increasingly automated. The value of engineering is moving toward deciding what to build, how to verify it and how to use increasingly capable systems to bring ideas into production faster.
CONCLUSION
The main change is not just faster coding but a redefinition of software engineering around delegation, verification and constant adaptation to rapidly improving models. As AI systems absorb more implementation work, human effort is shifting toward judgment, product direction and designing the loops that keep autonomous work reliable.
Explain this