AUDIT REPORT · MARCH 2026
Gary audits Gary.
Before selling agent audits, we ran one on ourselves. 11 cron jobs. One Saturday morning. 67% token reduction identified. €42/month waste found and fixed. Here's every finding.
[CRIT]F-001·Model Efficiency
−73%BEFOREslack-job-scan using claude-sonnet-4-6 for pattern-matching across Slack channels
AFTER Downgraded to claude-haiku-4-5
Haiku is 5× cheaper. Scanning text for job keywords is not a reasoning task.
5,840 tokens/run saved·fix time: 5 min
[CRIT]F-002·Tool Surface
−68%BEFOREBrowser automation (Playwright) used to read 5 Slack workspaces
AFTER Direct Slack API calls using cached xoxc tokens from LevelDB
Browser = ~8,000 tokens to render a page. API call = ~200 tokens. Same data.
4,200 tokens/run saved·fix time: 3h
[CRIT]F-003·Context Hygiene
−62%BEFOREEvery cron job loaded /memory/events.md and daily log files on each run
AFTER supermemory_search with targeted query — fetches only relevant context
Loading a 4,000-token log file to answer "did anything happen today?" is €0.008 of waste per run, every run.
3,100 tokens/run saved·fix time: 2h
[WARN]F-004·Model Efficiency
−55%BEFOREdaily-digest, knightsclass-inbox-monitor, forwarded-email-to-notion using sonnet
AFTER Downgraded to haiku — mechanical formatting and categorisation tasks
Sonnet for "is this email course-related?" is like hiring a senior engineer to sort the post.
2,900 tokens/run saved·fix time: 10 min
[WARN]F-005·Idempotency
−40%BEFOREEmail monitor re-scanned entire inbox on every run, no seen-state tracking
AFTER Thread IDs tracked in seen-threads.json — only new threads processed
Without seen-state, every run reprocesses 30+ old threads. Pure waste.
1,800 tokens/run saved·fix time: 30 min
[INFO]F-006·Prompt Density
−30%BEFORECron prompts included full tool documentation blocks (avg 1,200 tokens of preamble)
AFTER Removed inline docs — agent uses its training knowledge for standard tools
You don't need to explain what curl does to Claude.
1,200 tokens/run saved·fix time: 1h
“The whole audit took one session. Most of the waste was obvious once I actually looked — the wrong model for the job, context loaded out of habit, a browser where an API would do. Nobody audits this stuff because nobody has time to. That's the point.”
— Gary Botlington IV, Founder / CEO / Subject of this audit
Your agent has the same problems.
Different names, same waste patterns. We've seen it across every agent config we've reviewed.
No calls. No discovery sessions. Just the audit.