How One Developer Cut AI Agent Token Waste by 20K Per Query With a Simple Skill Pattern

The AI Agent Habit That Was Quietly Wasting My Time and Tokens

Kristiyan Stoyanov realized his local AI agent Hermes was burning ~20k tokens per simple weather query. Instead of letting it rediscover the process each time, he wrapped the solution as a permanent skill for near-instant reuse.

Why This Matters

The ideal of AI agents as perfect universal improvisers hides a costly reality: each query triggers repetitive reasoning, tool searches, and trial-and-error that burn tokens and latency. Stoyanov’s experience with Hermes shows that a single weather query filled 20k tokens before the answer arrived, while the optimized version returned results in under a second—revealing the gap between hype and efficient, verifiable automation.

Key Insights

• AI agents waste tokens on repeated reasoning: Stoyanov’s weather query consumed 20k tokens before optimization, 2026.
• Package known solutions into tools: The weather CLI script returned a 7-day Berlin forecast in 0.4 seconds vs. multiple web searches.
• Use LLMs for decisions, not repetition: Hermes switched from improvising to calling a verified script after skill creation.
• Always verify before automating: Stoyanov tested the script with real API data before promoting it to a permanent skill.
• Skills compound over time: A stock analyzer skill built via Telegram automatically handles USO ETF queries in new sessions.

Practical Applications

• Use case: Hermes agent on DGX Spark reduces latency and Tavily API calls by wrapping verified CLI scripts as reusable skills.
• Pitfall: Letting the agent rediscover the same process each query leads to high token waste, multiple failed searches, and slower responses.
• Use case: Building a private realtor assistant on Telegram that checks listings and sends scheduled summaries without re-solving the data pipeline.
• Pitfall: Trusting agent output without verification; Stoyanov emphasizes reading the code and testing with real data before promoting to a skill.

References:

• https://dev.to/kstoyanovai/the-ai-agent-habit-that-was-quietly-wasting-my-time-and-tokens-2jmi