Anthropic's Internal Scaling Playbook for Claude Code: Reducing Token Waste and Enabling Parallel Agents
Techniques from Anthropic's engineering team to run Claude Code efficiently at scale, focusing on system design, token optimization, and parallel execution to reduce operational costs.
Practical Summary
This entry breaks down a talk by Anthropic engineer Daisy Hollman, detailing how Anthropic's own teams configure and run Claude Code. It moves beyond simple prompting to building self-prompting systems, addressing key cost drivers like token bloat and enabling scalable workflows through features like Routines and Git worktrees.
Why It Matters
For teams deploying AI coding agents, operational costs are tied directly to token usage and workflow efficiency. Learning internal best practices from the model provider themselves offers a credible path to reducing waste, improving agent reliability, and scaling development capacity without proportional increases in cost or human oversight.
Shift from Prompting to System Design
The foundational principle is to build a system that prompts itself, rather than relying on manual prompt engineering for each task. This creates a more reliable and scalable foundation for cost-effective AI workflows.
Manage Token Bloat Proactively
Identify and eliminate unnecessary context that consumes input tokens ('token bloat') before sending a request. This is a direct cost-saving measure, as you pay per token. Review your system's context loading to ensure only relevant information is included.
Leverage the 'Routines' Feature
Explore and implement the 'Routines' feature within Claude Code. These are pre-defined sequences or templates that standardize common operations, reducing the need for complex, one-off prompts and improving consistency.
Implement Hands-Free Pipelines with /goal
Use the `/goal` command to initiate hands-free pipelines. This allows you to define a high-level objective, and the agent system manages the prompting and execution chain, enabling more autonomous and efficient task completion.
Scale with Git Worktrees for Parallel Execution
To run multiple agent instances simultaneously on different tasks without file conflicts, use Git worktrees. This allows each agent to operate on an isolated working copy of the repository, enabling parallel workflows that can dramatically increase throughput.