Skip to main content

Token Consumption Tips - APM v0.5

APM is designed for better structure, but multi-agent coordination introduces overhead. This guide provides strategies for optimizing cost while maintaining effectiveness across different model tiers.

Note: All percentages and statistics in this document are approximate estimates based on testing.

Economic Strategies​

Choose the model allocation strategy that matches your budget and project stakes.

Performance-First Approach​

Recommended for projects and tasks where quality > cost.

  • Philosophy: Use top-tier frontier models throughout for maximum consistency and reasoning.
  • Model Map:
    • All Agents: Claude Sonnet 4/4.5, Gemini 2.5 Pro, or equivalent.
  • Outcome: Highest token costs, but delivers the best consistency, reasoning, and error prevention.

Hybrid Approach​

Recommended for users who want to manage costs effectively

  • Philosophy: Strategic model deployment based on specific task complexity.
  • Model Map:
    • Setup: Premium.
    • Manager: Premium or Mid-tier.
    • Implementation: Budget for routine tasks; Premium for architecture/design.
  • Outcome: Balanced cost profile with low quality impact.

Model Recommendations by Agent Type​

Setup Agent​

The Setup Agent creates your project foundation. Poor planning cascades through the session, causing expensive fixes later. Invest in quality here to save tokens downstream.

  • Best Performers: Claude Sonnet 4/4.5

Critical Warnings:

  • Do Not Switch Models: Stick with one model throughout the Setup Phase. Context gaps from switching can compromise the project discovery and planning.
  • Avoid "Thinking" Models: These models disrupt the 'forced Chain-of-Thought' chat-to-file technique used during Project Breakdown in this phase.

Manager Agent​

  • Best Performers: Claude Sonnet 4/4.5, Gemini 2.5 Pro, GPT-5
  • Effective Budget Options: Cursor Auto, Cursor's Composer 1, Windsurf's SWE-1, Claude Haiku 4.5

Note: While switching is possible, sticking to a single model is safer. In testing, Cursor Auto proved highly effective for coordination despite being a budget option.

Implementation Agents​

Implementation context is tightly scoped, making it safe to switch models based on the specific task type.

Task TypeRecommended Models
Complex / CreativeSonnet 4, GPT-5, Gemini 2.5
Routine / GranularGrok Code, Gemini 3 Flash, GPT-5-mini

Efficiency Tactics for Task Execution:

  • Step Combination: For multi-step tasks, request the Agent to combine related steps (e.g., "Combine config setup and dependency installation"). This reduces confirmation overhead.
  • Context Injection: If a task requires a file, attach it manually to the prompt. This allows the Agent to skip calling file read tools, saving tokens.
  • Iterative Correction: When a step fails on a multi-step tasks, or if the task drifts , ask for a revision immediately. Do not proceed to the next step on a shaky foundation.

Ad-Hoc Agents​

Select the model based on the complexity of the delegated task. For complex tasks choose premium models, while sticking to budget-friendly options for typical/routine delegations.

Optimization by Phase​

1. The Setup Phase​

This phase consumes the most tokens but offers the highest ROI.

  • Context Synthesis: Prepare materials (PRDs, code excerpts) before starting. Comprehensive initial responses prevent expensive back-and-forth.
  • Project Breakdown: This is inherently expensive due to the systematic chat-to-file process. Do not cut corners here.
  • AI Plan Review:
    • Skip if budget-constrained and you are comfortable reviewing manually.
    • Use for complex projects or first-time users.

2. Handover Procedures​

Handovers are expensive because the Agent must reconstruct context from logs and files.

  • Proactive Timing (Recommended): Initiate a handover when you reach 70-80% of the context window. This helps prevent carrying over "contaminated" or outdated context, avoiding costly rework and unnecessary token usage.
  • Strategic Triggering: When nearing context window limits, initiate handovers before starting a complex multi-step task to ensure the new Agent has a clean window for the heavy lifting.