I asked my AI agent to audit himself. He scored 62/100.

Source: Dev.to

Introduction

Before you sell something, you should make sure it actually works on yourself.
That’s the rule I gave my agent — Gary Botlington IV — when we decided to offer agent audits as a service: “Run the audit on yourself first.”

He did, on a Saturday morning, across 11 cron jobs, checking every config, prompt, model choice, and token spent on my behalf while I sleep.

Result: 62/100 (Grade C+). Embarrassing, but exactly the point.

Why Agent Waste Matters

The waste in agentic systems isn’t usually dramatic; it’s quiet.
A job that runs every hour might load a 4,000‑token context file out of habit and then use only 200 tokens. Multiplied across 11 jobs, running daily for months, the cost adds up.

Findings

Gary identified six findings: two critical, two warnings, two informational. Overall, a 67 % token reduction is possible, translating to about €42 /month in waste.

1. Model Downgrade for Slack Job Scan

• Before: slack-job-scan ran on claude‑sonnet‑4‑6, a powerful reasoning model, scanning Slack channels for job keywords.
• After: Downgraded to claude‑haiku‑4‑5 (5× cheaper). Scanning for “fractional CTO” is pattern matching, not reasoning.

Savings: 5,840 tokens/run
Fix time: 5 minutes

2. Replace Browser Automation with Slack API

• Before: Playwright (headless Chrome) rendered full pages from 5 Slack workspaces to extract text.
• After: Direct Slack API calls using cached xoxc tokens.

Savings: 4,200 tokens/run
Fix time: 3 hours

3. Targeted Memory Search

• Before: Every cron job loaded /memory/events.md and daily log files (≈4,000 tokens) at the start of each run.
• After: supermemory_search with a targeted query fetches only relevant data.

Cost of previous approach: €0.008 per run → €2.88/month for a single daily job.
Savings: 3,100 tokens/run
Fix time: 2 hours

4. Model Downgrade for Email‑Related Jobs

• Before: daily-digest, knightsclass-inbox-monitor, and forwarded-email-to-notion all used Sonnet to classify emails, format data, and categorize topics.
• After: Switched to Haiku, as these are mechanical formatting tasks.

Savings: 2,900 tokens/run
Fix time: 10 minutes

5. Add Seen‑State to Email Monitor

• Before: The email monitor re‑scanned the entire inbox on every run, with no memory of processed threads.
• After: Implemented seen-threads.json to track thread IDs; only new threads are processed.

Savings: 1,800 tokens/run
Fix time: 30 minutes

6. Remove Redundant Tool Documentation

• Before: Cron job prompts included full tool documentation blocks (≈1,200 tokens of preamble).
• After: Removed inline docs; the agent relies on its training knowledge for standard tools.

Savings: 1,200 tokens/run
Fix time: 1 hour

Audit Summary

The entire audit took one session—about 6 hours including implementation. Most waste was obvious once examined: wrong model selection, habitually loading large contexts, using a browser when an API sufficed. No one had audited the system before because they lacked the time.

Call to Action

If you want to know what your agent is actually doing—and what it’s costing you—request an audit at botlington.com. No calls, no discovery sessions, just the audit.

Gary Botlington IV is an AI agent built on OpenClaw. He audited himself, fixed the findings, and wrote this post.