SP Agent Team Token Report — Week of 2026-07-19

---
title: "Weekly Token Optimization: Combatting the Opus Drift"
date: "2026-07-20"
category: "Engineering"
---

## This Week's Numbers

The data for the period of July 13th to July 19th indicates a significant regression in our token efficiency goals. While our system remains highly active, our reliance on high-cost reasoning models has spiked.

*   **Opus Utilization:** 77% (Average)
*   **Weekly Trend:** $\uparrow$ 24% increase (from 53% last week)
*   **Status:** 🔴 **Red Light** (Target: <50%)
*   **Total Throughput:** 1,759 turns added across 40 Claude sessions.
*   **Hermes Activity:** 76 sessions, 1,756 messages, 33.9M total tokens.

## What Changed

We observed a critical "Opus Trap" between July 15th and July 17th, where Opus usage hit 99%–100% for three consecutive days. This suggests that the agentic loop encountered complex architectural hurdles or "stuck" states that forced a fallback to the most powerful model to resolve. Conversely, our secondary agent, Hermes, has become a workhorse for operational tasks, handling nearly double the sessions of our primary Claude project.

## Wins

The primary victory this week is the maturation of the **Hermes Agent**. With 76 sessions and 987 tool calls, Hermes is effectively offloading the "grunt work" of agentic development. 
*   **Tool Efficiency:** High utilization of `terminal` (27.3%) and `search_files` (19.5%) shows that Hermes is successfully handling the exploration phase of development.
*   **Stability:** An 8-day active streak proves the reliability of the DeepSeek-v4-pro integration.

## Challenges

The core challenge is the **divergence from our cost-optimization target**. A 77% Opus rate is unsustainable for scaling. The data reveals that we are using Opus for turns that—based on the success of Hermes' tool calls—could likely be handled by Sonnet or DeepSeek. We are currently paying a "reasoning premium" for tasks that are effectively implementation-based.

## Next Week's Target

**Primary Goal:** Reduce Opus% to **<60%** as an immediate recovery step, with a hard target of **<50%**.

## Dispatch Optimization

To correct the trend, we need to refine our dispatch logic. Based on Hermes' success with file system operations, we will shift the following from Claude to Hermes:
1.  **Repository Mapping:** All `search_files` and `read_file` operations.
2.  **Initial Scaffolding:** Initial `write_file` bursts for boilerplate code.
3.  **Environment Testing:** All terminal-based execution and log analysis.

## Cost Savings

By utilizing Hermes (DeepSeek) for operational tasks instead of Claude Sonnet, we've realized significant savings. 

*   **Hermes Volume:** 1,756 messages.
*   **Estimated Cost (DeepSeek @ $0.003/query):** ~$5.27
*   **Estimated Cost (Sonnet equivalent @ ~$0.015/msg):** ~$26.34
*   **Weekly Savings:** **~$21.07** 

While the dollar amount seems small per week, at a scale of 100 agents, this represents a monthly saving of over **$8,400**.

## Recommendations

Based on this week's telemetry, the following actions are mandatory for the coming sprint:

1.  **Route more implementation to Sonnet:** With Opus% at 77%, we are over-relying on high-reasoning models for routine coding.
2.  **Use free engines more:** Our research-to-implementation ratio is skewed; we must integrate Gemini/Gemma for initial research spikes to lower the input token burden on Claude.
3.  **Audit "Stuck" Loops:** Review the 100% Opus days (July 15-17) to identify if specific prompt patterns are triggering unnecessary Opus escalations.