Why can a multi-agent design cost more than expected?

Each agent adds more than a single model call. It may need its own prompt, copied context, tool output, generated response, retry path and handoff message. The total cost is the sum across the whole workflow, including intermediate artefacts that users never see.

Does adding more agents usually improve answer quality?

No. More agents can duplicate work, share the same flawed premise, or obscure where an error entered the workflow. Quality improves only when the agents have meaningfully different responsibilities, constrained interfaces, and a reliable way to judge whether the combined result is better than a simpler baseline.

What should be logged in a multi-agent system?

Log the prompts, model outputs, tool calls, routing choices, retries, handoff data and final selection logic. Without those records, debugging becomes guesswork because the final result may depend on several hidden intermediate decisions rather than one obvious failure point.

Multi-agent orchestration

07.05 · Concept

The cost of a committee

Measure what multi-agent costs in tokens, latency and debuggability before committing.

Multi-agent orchestration is a production fan-out of model calls, not free extra reasoning. Its cost comes from accumulated prompts, context, tool results, retries and intermediate outputs, while latency follows the critical path. A committee is only worth considering when roles, handoffs, evaluation and logging justify the added operational complexity.

What this lesson answers

how to estimate multi-agent token cost
when are multiple AI agents worth it
why multi-agent systems increase latency

Notes

A multi-agent system is not free intelligence; it is a distributed system made of model calls. Every extra agent usually means extra prompts, extra context passed around, extra tool calls, extra retries, and extra intermediate outputs. Before using a committee, estimate the full call graph: how many model invocations happen in the common case, how large each input and output is, and which calls can run in parallel versus sequentially.

The practical mental model is a production request fan-out. A single-agent design is one service call with one trace.

Common questions

Why can a multi-agent design cost more than expected?: Each agent adds more than a single model call. It may need its own prompt, copied context, tool output, generated response, retry path and handoff message. The total cost is the sum across the whole workflow, including intermediate artefacts that users never see.
Does adding more agents usually improve answer quality?: No. More agents can duplicate work, share the same flawed premise, or obscure where an error entered the workflow. Quality improves only when the agents have meaningfully different responsibilities, constrained interfaces, and a reliable way to judge whether the combined result is better than a simpler baseline.
What should be logged in a multi-agent system?: Log the prompts, model outputs, tool calls, routing choices, retries, handoff data and final selection logic. Without those records, debugging becomes guesswork because the final result may depend on several hidden intermediate decisions rather than one obvious failure point.

Short definition: what is cost of a committee?

07.04 · Multi-agent orchestrationOrchestrator and workers 08.01 · Permissions and sandboxingLeast privilege, for a loop