4× or 15×: topology decides the agent bill, cadence decides how often you pay it

Agent cost usually gets discussed per run, which hides the thing that actually determines the bill. In Anthropic's data, agents typically use about 4× more tokens than chat interactions, and multi-agent systems use about 15× more tokens than chats. Those are multipliers, and they apply to whatever cadence has been chosen. A system that runs when someone asks it to pays once. A system that runs unsupervised every day pays the multiplier every day.

The multiplier lands on the schedule

Before pricing an agent, write down how often it runs and how many outputs each run produces. One platform publishes eight SEO articles every day in seven languages across four country domains, with manual intervention required: 0. Another publishes articles autonomously on a daily basis. A cadence like that turns a topology choice into a recurring cost, so the question is not whether an agent could do the work but which part of the work needs one.

A workflow is the cheaper default

Anthropic draws a clean line. Workflows are systems where LLMs and tools are orchestrated through predefined code paths, while agents are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks. A daily publishing target contains a lot of predefined path. The number of pieces, the languages, the destinations and the order of steps are fixed before the run starts. Encode those, and no model has to rediscover them.

Spend agency where the path is unknown

Dynamic direction earns its multiplier where the next step genuinely depends on what came back. Topic selection driven by search trends is one of those places: the input changes daily and the branch cannot be written in advance. Keep that scope narrow and let the surrounding structure stay deterministic. The autonomous nature of agents means higher costs and the potential for compounding errors, so each grant of discretion should be somewhere hard-coding was not an option.

Deterministic recovery beats improvised recovery

Recovery is a common excuse for extra agency and rarely needs it. One autonomous publishing platform runs with three model fallback levels and weekly content operations time of about zero hours. A ranked list of fallbacks is a predefined path, not something a model has to reason its way through. Anthropic describes combining the adaptability of AI agents with deterministic safeguards like retry logic and regular checkpoints, alongside systems that resume from where the agent was when errors occurred.

Caps belong outside the model

Stopping conditions, such as a maximum number of iterations, keep an agent under control, and the same condition caps its spend. Rate-limiting reduces the number of undesirable actions that can take place within a given time period, which is cost control as much as safety control. Put both in the orchestrator rather than the prompt, and do extensive testing in sandboxed environments at the real cadence. A loop that costs little once costs differently at a daily rhythm.

Memory is where the multiplier stops growing

Context grows across a long run, and every added token is billed at the topology's multiplier. Anthropic implemented patterns where agents summarise completed work phases and store essential information in external memory before proceeding to new tasks. That keeps the live context bounded and gives a checkpoint something to resume from. One write serves two purposes: a smaller prompt for the next phase, and a restart point that avoids replaying everything already finished.

Price the agent as cadence times topology, not as a single run. Count the runs per day, mark the steps that truly need a model to choose what happens next, and move the rest into predefined code paths with iteration caps and rate limits enforced outside the model. Then checkpoint completed phases into external memory, so a long run stays bounded and a failure costs one phase instead of the whole thing.

Sources

Anthropic — Building effective agents — https://www.anthropic.com/engineering/building-effective-agents

Anthropic — How we built our multi-agent research system — https://www.anthropic.com/engineering/multi-agent-research-system

OWASP Gen AI Security Project — LLM06:2025 Excessive Agency — https://genai.owasp.org/llmrisk/llm062025-excessive-agency/

Neurolinks case study — Four markets, one codebase — https://neurolinks.be/work/tatano-energy

Neurolinks case study — A personal brand that publishes itself — https://neurolinks.be/work/matthieu-pesesse-media