The Multi-Agent Workflow Token Budget Planner estimates token demand across several cooperating agents rather than treating the workflow as one model call. It accounts for agent count, turns per agent, shared context, agent-specific context, generated output, coordination messages, and rerun overhead.
This is useful for supervisor-worker, debate, review, or specialist-agent systems where the same shared state may be repeated many times. The breakdown highlights how orchestration and duplicated context increase the budget beyond the visible final answer.
1. Set workflow volume
Enter the number of full multi-agent runs in the planning period.
2. Describe topology
Enter participating agents and average turns completed by each agent.
3. Estimate repeated context
Separate shared context from agent-specific context for one turn.
4. Add generated and coordination text
Include expected output per turn and orchestration messages per workflow.
5. Allow for reruns
Enter the share of full workflows expected to repeat and review the final budget.
The total represents all estimated tokens consumed across every agent turn and coordination step in the workflow batch.
If agents pass growing transcripts to each other, use average context per turn or model each phase separately because context may not remain constant.
Given: 500 runs, 3 agents, 4 turns each, 900 shared-context tokens, 500 agent-context tokens, 300 output tokens, 700 coordination tokens, and 8% reruns.
Calculation: Turn tokens = 900 + 500 + 300 = 1,700. Base workflow = 3 × 4 × 1,700 + 700 = 21,100 tokens. Adjusted workflow = 21,100 × 1.08 = 22,788. Total = 22,788 × 500 = 11,394,000 tokens.
Result: Plan for about 11.39 million tokens.
Yes, when it makes model calls using the same turn assumptions. Otherwise include its tokens in coordination overhead.
Calculate each agent role separately and sum the token totals.
It covers routing prompts, summaries, handoffs, voting, or synthesis that is not already included in agent turns.
Only through the average values you enter. Growing transcripts should be modeled with phase-specific averages.
It assumes a full-workflow equivalent. Convert partial retries into an equivalent percentage of the full base token budget.