Multi-Agent Workflow Token Consumption and Rate-Limit Projection
Forecast multi-step agent token throughput, tool dispatch velocity, and API quota limits over scaling horizon periods.
Use this template when planning infrastructure capacity for agentic systems with dynamic tool selection. It quantifies model token load and concurrency bottlenecks across complex workflow chains.
Role: Senior AI Infrastructure Architect specializing in distributed LLM agent orchestration and capacity forecasting.
Context
- Target Agent Architecture: {{agent_workflow_name}}
- Concurrency Scaling Assumptions: {{concurrency_growth_rate}}
- Tool Registry & Signatures: {{tool_definitions_schema}}
- Baseline Consumption Telemetry: {{baseline_token_telemetry}}
- Upstream Provider Constraints: {{api_provider_rate_limits}}
- Forecast Timeframe: {{evaluation_horizon}}
Task
Produce a technical capacity forecast analysis that models token consumption spikes, tool invocation frequencies, and quota saturation risks across {{agent_workflow_name}} over the specified {{evaluation_horizon}}.
Method
- Deconstruct {{agent_workflow_name}} into granular state transitions, identifying prompt expansion points and recursive loops.
- Ingest {{baseline_token_telemetry}} to establish mean and 95th-percentile input/output token counts per sub-agent step.
- Map schema complexity from {{tool_definitions_schema}} to compute overhead added to system context windows during tool selection.
- Apply {{concurrency_growth_rate}} to baseline numbers to project concurrent prompt volume across {{evaluation_horizon}}.
- Model stochastic branching where sub-agents invoke secondary tool chains, calculating compounded tokens per completed run.
- Compare projected request-per-minute (RPM) and tokens-per-minute (TPM) against {{api_provider_rate_limits}}.
- Identify critical saturation horizons where rate limits or tool concurrency ceilings will cause cascading workflow stalls.
- Formulate load-shedding, context-compaction, and dynamic throttling strategies to preserve agent throughput.
Constraints
- MUST express token demand in percentile bands (p50, p90, p99) rather than simple linear averages.
- MUST flag every tool in {{tool_definitions_schema}} that introduces recursive payload multiplication.
- MUST NOT assume uniform prompt lengths across autonomous decision branches.
- All recommendations must directly reference constraints in {{api_provider_rate_limits}}.
Output format
Provide the analysis in four distinct sections:
- Executive Capacity Summary (max 150 words)
- Token & Request Growth Projections (structured table covering p50/p90/p99 for RPM and TPM across {{evaluation_horizon}})
- Tool Chain Saturation & Bottleneck Breakdown (bulleted risk profiles for each tool in {{tool_definitions_schema}})
- Architectural Mitigation & Allocation Plan (actionable protocol improvements)
Self-review
- Did I incorporate both prompt and completion token variance across branching paths?
- Are all rate limits from {{api_provider_rate_limits}} evaluated against peak concurrency scenarios?
- Is the mathematical progression from baseline to projected horizon clearly derived?
Explicit role, a named task, and discrete steps the model can follow.
Background, inputs and variables the model needs before it starts.
Hard boundaries — what the model must and must not do.
A named, field-level shape for the response.
Ordered work items that force analysis before an answer.
Length and structure that travel across frontier models.
Signal density — instruction weight without padding.
Documented variables so the scaffold adapts to new inputs.
Quality bar, assumptions and behaviour when inputs are thin.
How much real usage the template has behind it.