Claude Tag now lands 65% of product engineering PRs for Anthropic’s Claude Code team

Anthropic’s internal deployment of Claude Tag, a collaborative coding agent integrated with Slack, now automates 65% of product engineering pull requests for the Claude Code team. This milestone reflects a broader shift toward proactive, multiplayer coding agents that operate within team collaboration tools rather than as isolated developer utilities.

The team attributes this adoption to Claude Tag’s multiplayer-by-default design, which allows engineers and non-engineers to collaborate on pull requests, and its proactive capabilities, such as autonomously monitoring bug reports and proposing fixes. Additionally, team memory enables the agent to retain preferences and context shared in natural language within Slack channels, reducing repetitive instructions and improving consistency.

Claude Code’s system prompt has undergone an 80% reduction in size with newer models like Fable 5 and Opus 4.8. The team found that adding examples to system prompts and including lists of hard constraints (e.g., "don’t do X and don’t do Y") often degraded model performance. Instead, they now prioritize providing context and fewer, softer constraints, allowing the models to exercise judgment. This change is only feasible with frontier models capable of handling more nuanced instructions.

Code review at Anthropic has evolved to rely increasingly on automated systems, particularly for changes in the "outer layers" of the codebase. Critical components still undergo manual review by designated code owners, but the team has built confidence in automated review through iterative improvements and a growing suite of evals. Incidents are analyzed to update code review policies, ensuring automated systems catch regressions. This approach has enabled the team to reduce human involvement in routine reviews while maintaining quality standards.

Auto mode, a feature that enables safe, long-running operations for coding agents, is central to Anthropic’s security model. The team has conducted thousands of evals and commissioned red-team exercises to test auto mode against adversarial inputs, including prompt injection and data exfiltration risks. Anthropic claims auto mode mitigates the main categories of risks more effectively than average human reviewers, though it does not guarantee 100% coverage. This feature is foundational to Claude Tag’s ability to operate securely in shared Slack channels.

The team also highlighted the cultural shift within Anthropic, where "ant fooding"—internal dogfooding—drives product decisions. Features are first deployed to employees, and only those demonstrating strong user retention are released externally. This approach ensures that new capabilities, such as proactive bug fixing and team memory, are validated against real-world usage before broader rollout.