Anthropic provides its user manual for scaling agents: layers, tokens, reliable execution

In plain terms Two engineers from Anthropic - Angela Jiang and Katelyn Lesse - publish the lessons learned in industrializing AI agents for their clients. The message: it's not the model that limits scalability, it's the surrounding architecture. Three pillars: breaking down the stack into layers, maintaining a strict token budget, and ensuring execution.

The context

The dominant discourse in 2025 was: "the models will get there." Anthropic's experience shifts the focus in 2026: on long agent workflows, reliability comes from the harness, not the model. This aligns with the thread we follow: the cost per token slides into the opex, Adam Mosseri mentions token ceilings per engineer (Meta), and AI licenses are cut at Uber and Microsoft when the budget spirals out of control.

The data

• Primary source: publication by Angela Jiang and Katelyn Lesse (Anthropic), reported by Tech in Asia on 15/07/26.
• Key message reported: AI scalability in enterprises relies on a layered architecture, a token strategy, and reliable execution—not just better models.

Under the hood

The "layered architecture" pattern, in practice, separates three responsibilities:

1. Orchestration - decides which agent runs, with which token budget and which deadline. This is the layer that refuses a run if the budget is exceeded.
2. Execution - executes each tool-call with retries, timeouts, output validation, and idempotence. Without this, an agent that "works" in a demo becomes an incident in production.
3. State - persists what needs to survive between turns: the artifacts and decisions, not the entire conversation context. Context costs money, so it is rationed.

The token budget is the structuring constraint: if an agent can burn 500k tokens per run and runs 1000 times a day, the cost quickly exceeds a junior engineer's salary. The budget is therefore set at the architectural level—not at the prompt level.

Scenarios & Risks

• Short term: teams building agents without token caps will be caught up by the CFO (the token-budget-caps thread already documents cases from Uber, Microsoft, Mosseri).
• Product risk: moving to reliable execution requires instrumentation—traces, replays, cost per run—which most demo agents ignore. Many POCs will not survive this instrumentation cost.

So what

For a lead engineer driving an agent project in 2026: the first priority is no longer prompt engineering. It's the budget and the trace. Instrument the cost per run before adding capabilities. Otherwise, the conversation with finance will arrive at the worst time—that moment when the agent finally starts working.