SP Agent Team Token Report — Week of 2026-06-21

---
title: "Weekly Token Optimization Report: The Opus Spike"
date: "2026-06-22"
category: "Engineering"
---

## This Week's Numbers

The primary objective for the SuperPortia agentic system is to maintain an **Opus% of <50%** to balance reasoning quality with operational costs. This week, we hit a significant roadblock.

- **Average Opus Usage**: 78% (Last week: 57%)
- **Trend**: $\uparrow$ 21% worsening
- **Status**: 🔴 **RED LIGHT**
- **Claude Volume**: 1,445 total turns (1,139 Opus / 306 Sonnet)
- **Hermes Volume**: 123 sessions / 201.8M tokens processed

## What Changed

We observed a drastic surge in Opus dependency between June 16th and 18th, peaking at **95% on Tuesday**. This suggests that several high-complexity architectural tasks were routed to the heaviest model by default. While Sunday (June 21) showed a promising drop to **34% Opus usage**, the weekly average remains far above our efficiency threshold.

## Wins

Despite the Claude imbalance, the **Hermes Agent** (our secondary dispatch layer) is performing exceptionally well as a high-volume "workhorse." 
- **Massive Throughput**: Hermes handled over **201 million tokens**, primarily utilizing `gemini-3.5-flash` for bulk processing.
- **Tool Proficiency**: The agent is effectively utilizing the `terminal` (46.4%) and `read_file` (24.2%) tools, automating the "grunt work" of codebase exploration without consuming expensive Claude tokens.
- **Consistency**: A perfect 8-day active streak, proving Hermes is now a stable part of the daily workflow.

## Challenges

The "Red Light" status is a direct result of **Model Drift**. We are defaulting to Opus for tasks that Sonnet 3.5 could likely handle. The data shows that when we lean into Sonnet (as seen on June 21), the system remains functional, yet the inertia during the work-week pulls us back toward Opus.

## Next Week's Target

- **Opus% Target**: < 50%
- **Primary Goal**: Force a shift of implementation-heavy tasks from Opus to Sonnet.

## Dispatch Optimization

To correct the Opus% trend, we will shift the following logic from the primary Claude dispatcher to Hermes:
1. **Log Analysis & Monitoring**: Move all `daily-monitoring-report` and `llm-usage-report` logic entirely to Hermes.
2. **Initial Codebase Mapping**: All `search_files` and `read_file` discovery phases must be completed by Hermes before escalating to Claude.
3. **Routine Cron Tasks**: Shift the 78 sessions currently running on `cron` to be exclusively managed by the Flash-tier models in Hermes.

## Cost Savings

By offloading high-token, low-reasoning tasks to Hermes (DeepSeek/Gemini), we avoid the "Sonnet Tax." 

Assuming the 4,323 messages handled by Hermes were instead routed to Claude Sonnet (est. avg $0.015/turn) vs. the DeepSeek Flash rate (~$0.003/query):
- **Estimated Sonnet Cost**: $64.85
- **Actual Hermes Cost**: $12.97
- **Weekly Savings**: **~$51.88** (on a small scale; this scales linearly as we add more agents).

## Recommendations

Based on this week's telemetry, we are implementing the following actions:

1. **Route more implementation to Sonnet**: With Opus usage at 78%, we are over-relying on the heaviest model. We will update the routing prompt to prefer Sonnet for all code edits.
2. **Review rate too low — check enforcement hooks**: We recorded **0 L3 read-explore fires** despite 222 new files. This indicates our automated review triggers are not firing; we need to audit the `evaluate.sh` hooks.
3. **Optimize Peak Hour Dispatch**: Activity peaks at 9 AM. We will implement a "Morning Warm-up" mode where Hermes performs all initial indexing to prevent the 9 AM Opus spike.