Model the four pressures called out in the source conversation: a larger context window, more subagents, inherited full history, and deeper reasoning. Then compare a leaner run before spending the tokens.
Changes recalculate immediately and persist locally.
How do inherited context, subagent counts, and reasoning multipliers compound total token consumption across agent work rounds?
This planner decomposes multi-agent LLM token consumption into primary coordinator history, inherited subagent context, and reasoning plus output generation. When subagents receive full conversational history on every round, context reads scale multiplicatively with subagent count and round count. Pruning or summarizing inherited context and capping active history with compaction limits reduces the compounding growth rate before API requests are dispatched.
This calculator applies deterministic heuristic formulas rather than live provider telemetry. It does not account for provider prefix prompt caching, selective retrieval-augmented prompt assembly (RAG), speculative decoding, tool call schema overhead, or variable pricing tiers across distinct foundation models.
Click the 'High-burn preset' button. The context window expands to 400k at 70% active history (capped by compaction at 280,000 tokens), subagents increase to 8, reasoning rises to high (1.8x), and rounds double to 6. Inherited context expands to 13,440,000 tokens, elevating the planning total to 15,452,640 tokens (8.26x the baseline reference).
Active history is calculated as the minimum of the context window multiplied by the history usage percentage and the compaction ceiling. Primary coordinator history equals active history multiplied by total work rounds. Inherited context scales as active history multiplied by the inheritance ratio, the number of subagents, and the work rounds. Output and reasoning tokens equal the base output per round multiplied by rounds and scaled by one plus the product of subagent count and reasoning depth multiplier.
Because inherited context compounds across subagents and execution rounds, passing clean task briefs (0.05 inheritance factor) or compact summaries (0.28 factor) instead of full conversational history (1.0 factor) substantially curtails token velocity. Combined with aggressive history compaction limits, multi-agent pipelines avoid re-reading redundant historical context on every iterative call.
This planning tool models token usage rather than observing an account. Default context window is two hundred seventy-two thousand and history use is forty-five percent. Their product is one hundred twenty-two thousand four hundred. Active history takes the minimum of that product and the compaction threshold, two hundred thirty-three thousand. At defaults the cap does not bind. Compaction is represented by a cap, not an actual summarization process. Across three rounds, primary history contributes three hundred sixty-seven thousand two hundred tokens. Four subagents inheriting the entire active history contribute one million four hundred sixty-eight thousand eight hundred. The output and reasoning term adds thirty-four thousand five hundred sixty, using eighteen hundred output tokens times three rounds times one plus four times one point three five. The total is one million eight hundred seventy thousand five hundred sixty. The displayed lean comparison changes multiple inputs: context is capped at two hundred thousand, history at thirty percent, subagents at two, inheritance becomes point two eight, and reasoning becomes one. With the default three rounds and eighteen hundred output tokens, lean terms are one hundred eighty thousand, one hundred thousand eight hundred and sixteen thousand two hundred. Total is two hundred ninety-seven thousand, about eighty-four percent lower. That is a hypothetical comparison, not measured savings or a causal test of inheritance alone.