Anthropic explains the strict token budgets required to scale

This article summarizes an episode of Sequoia Capital’s video series featuring Anthropic’s Angela Jiang and Katelyn Lesse.

Anthropic’s Angela Jiang and Katelyn Lesse / Photo credit: Sequoia capital

As companies move from AI demos to agents that touch real work, the costly question is how much control they have over each step. Angela Jiang, head of product for Anthropic’s Claude Developer Platform, and Katelyn Lesse, head of engineering for the platform, believe that progress depends on breaking complex AI jobs into smaller pieces.

By designing platforms with distinct coordination layers and giving tokens specific jobs, companies can make AI choices based on clear costs, safety, and performance.

AI platforms are getting more organized

Jiang argues that business AI systems should separate information, action, and planning into different areas. Doing so makes it easier to blend AI assistants with existing office tools while giving companies greater control over reliability and costs.

Her view organizes the system by the responsibility each part carries:

• At the information level, the AI receives instructions, tools, skills, memory, and background details so it can function across different products.
• Managed computer systems give long-running AI assistants a safe space to work. Here, they can remember their progress, manage file access, control costs, and safely pause and resume tasks.
• Basic operating tools become the foundation for longer jobs. Newer AI models need less manual guidance and are naturally better at finishing tasks, tracking progress, and fixing errors.
• Planning levels combine tools, memory, and fact-checking when a project requires asking the AI for help multiple times.
• A company should focus on the part that makes their business unique. New AI startups might want basic building blocks, while large businesses often prefer reliability, rule compliance, and easy setup.

The plan moves forward as AI assistants take on longer tasks and need a strategy for work that happens over time.

Jiang explains, “At the coordination layer, we’re beginning to think about things called strategies. It’s almost like a meta-harness… We’ll move more and more from the knowledge layer to the execution layer, and from the execution layer to the coordination layer.”

Because of this, basic access to the AI becomes the least unique part of buying the software. The harder question is whether the system can remember background details, keep security rules in place, recover from mistakes, and leave a record the company can check.

The same logic explains Anthropic’s decision to offer the same basic tools to both their own team and outside developers. This reduces the risk of creating tools that only solve specific internal problems.

Token budgets need clear operating roles

Once systems are organized into layers, tokens become a resource that teams plan out carefully instead of a single pool of computing power.

Jiang and Lesse argue that organizations should assign tokens to specific roles and measure if each step improves accuracy, speed, confidence, or safety enough to justify the cost.

In that view, a token budget has distinct jobs inside the AI process:

The real value is behind the scenes

Mixing open systems with specialized tools