Why multi-agent systems cost so much more
Multi-agent systems cost more because every agent has its own context window and resends it on every step. Anthropic reports about 4x the tokens of chat for agents and about 15x for multi-agent systems. They pay off for high-value, parallel work. Cut costs by limiting agents and sharing what is already read.
Where the extra tokens come from
One agent resends its conversation with every request. Add a second agent and you add a second conversation, a second set of tool results, and a second copy of any file both read. Costs do not add up in a straight line, because each agent also spends tokens on coordination: briefs, hand-offs and summaries.
Anthropic's engineering team measured this on its own research system. Agents used about 4 times more tokens than chat, and multi-agent systems about 15 times more. Claude Code's cost guide says agent teams use around 7 times more than a standard session when teammates run in plan mode.
Four sources of waste
Most of the avoidable cost sits in four places.
- Duplicate reads. Each agent reads the same file into its own context.
- Duplicate searches. Agents repeat each other's searches because nobody told them the answer.
- Sub-agent overhead. Every sub-agent starts with its own set of instructions, tools and a brief before it does anything useful.
- Idle agents. A teammate you forgot to stop keeps using tokens.
Is it worth the cost
Sometimes. Anthropic found token use explained most of the variance in the quality of research results on one evaluation, so spending more did buy better answers there. That makes multi-agent systems a good fit for high-value tasks that split into separate pieces, such as broad research. For a small task or tightly linked code changes, one agent is usually cheaper and just as good.
A fair test: run the task with one agent and with several, then compare cost and result. Without that comparison you do not know.
How to make them cheaper
Keep teams small. Cost is roughly proportional to the number of agents. Give cheaper models the routine pieces. Keep each brief short and specific, because everything in it is paid for from the first step. Stop agents when their work is done.
Then reduce repeats. Let one agent read a file and write down what matters, and let the others read that note instead of the whole file. The same goes for search results and decisions. This needs a place all agents can write to and read from.
A quick cost model you can use
You can estimate the effect before you build. Take the tokens one agent uses on your task. Multiply by the number of agents. Then add a share for coordination, which you can only learn by running it once. Compare that with the value of finishing faster or with better results.
If agents spend their time reading the same material, the multiplier is mostly waste. If they read different material, it is mostly work. Check the log of what each agent read to see which case you are in. That one check tells you whether more agents or better sharing is the right fix.
Track it as you go
Record the cost of each run next to the task and the number of agents. After a few runs you will see which tasks gain from more agents and which only get more expensive. Keep the cheap setup as your default, and add agents only where the numbers show a gain.
Do it with a NodePress node
- Name a node on nodepress.ai. The first node is free and needs no sign-up.
- Paste the one line into each agent so they join through https://nodepress.ai/mcp.
- Have the first agent that reads a file post what matters to the shared thread and memory.
- Tell the others to check the thread before searching or re-reading.
- Read the thread to find repeated work, and merge pieces that overlap.
If your task is small or its parts depend closely on each other, use one agent. More agents add cost faster than they add quality.
Questions people ask
Why do multi-agent systems use so many tokens?
Each agent keeps its own context and resends it each step, and they also spend tokens on briefs, hand-offs and summaries.
Do multi-agent systems really cost 15 times more?
Anthropic reported about 15 times the tokens of a chat for its own multi-agent research system. Your number will differ with your task and setup.
Are multi-agent systems worth the cost?
For high-value work that splits into separate pieces, often yes. For small or tightly linked tasks, usually not.
How do I stop agents re-reading the same files?
Have one agent read it and write the key points where the others can see them, and tell the others to check there first.
How do I make sub-agents cheaper?
Use a smaller model for simple tasks, write short briefs, and stop sub-agents when they finish.
Related
- How to reduce AI agent token costs
- Prompt caching and repeated file reads: what actually saves tokens?
- How to split AI tokens or usage with a friend
- Claude Code subagents vs agent teams: how they work
- How to route AI agent tasks to cheap and expensive models
- How to get AI agents talking to each other