$1 a day. That is what my AI agent costs to run after two weeks of production use.
It generates invoices, scans three email accounts for receipts, reconciles bank statements, and delivers a morning capacity brief based on my Garmin sleep and training data. If something breaks, it files a Linear ticket. I come back with Claude Code and fix the issue.
The agent doing all this does not reason about what to do next. It calls tools on a cron schedule. I do not rely on its thinking. That is the whole point.
What $80 Overnight Taught Me
I tried an autonomous AI agent earlier this year. Gave it a task, went to sleep, woke up to an $80 bill and zero output. The agent spent the night browsing movie websites in loops without saving anything or producing results.
The model was cheap — DeepSeek-tier pricing, fractions of a cent per token. The architecture was the problem: no definition of "done," no cost ceiling, no schedule. Just permission to act and an API key.
88% of AI agent projects fail before reaching production. Most fail for the same reason — autonomy without structure.
The fix I proposed then was boring: scheduled triggers, approval gates, summary reports. The same things that make n8n workflows reliable. So I tested that thesis with an agent called Hermes.
What Hermes Actually Does
Hermes runs on a Raspberry Pi 5 in my apartment. Three Docker containers — the agent gateway built on Nous Research's Hermes Agent framework, a cache proxy for prompt caching, and a Google Suite MCP server. Eighteen skills. Nine scheduled tasks. Accessible through Telegram.
After killing my own agentic runtime earlier this year, I stopped building infrastructure and started configuring it. Hermes is someone else's framework with my skills bolted on.
Two tasks deliver most of the value:
Month-end close. On the first of every month, Hermes generates sales invoices for my clients as Lithuanian VAT-compliant PDFs. It scans three Gmail accounts for purchase invoices, classifies them — ready, portal-only, already submitted — and forwards confirmed documents to my accountant's inbox. Then it reconciles my bank statement against the submissions. This used to take me a couple of hours. Now I review a report.
Morning capacity brief. Every morning, Hermes pulls my Garmin data — sleep score, body battery, resting heart rate, stress — and delivers a GREEN, YELLOW, or RED assessment through Telegram. I know whether to push hard or take it easy. No opening the app, no interpreting numbers.
The key decision: Hermes is a dynamic tool caller, not a reasoning engine. It does not decide what to do. The cron schedule decides. It does not interpret results. The MCP tools return structured data. Hermes is glue between a schedule and a set of functions.
How Much Does an AI Agent Cost to Run
Running a personal AI agent costs $1 per day at steady state. My setup processed 155 million prompt tokens over two weeks through Qwen 3.6 Plus on OpenRouter at $0.30 per million input tokens. A cache proxy saves more than the total bill — $34 in cache savings against $32 spent. No frontier model needed.
The total two-week cost was $32, but $21 of that burned in the first three days before the cache proxy worked. Once caching kicked in, daily cost dropped to about $1. The cache proxy injects ephemeral breakpoints on text blocks over 4,000 characters — Qwen models need explicit markers, unlike OpenAI which caches implicitly. One sidecar container cut my bill in half.
For context, managed AI agent platforms charge $300 to $800 per month for a single-purpose agent. Enterprise multi-agent systems run $1,000 to $3,000 per month. Running your own on a cheap model with caching is a different category.
The difference between autonomous and scheduled AI agents is not the model or the price per token — both Clawbot and Hermes used cheap models. The difference is structure. An autonomous agent decides what to do, when to act, and when to stop. A scheduled agent runs defined tasks on a cron schedule and stops when the task completes. As of mid-2026, the scheduled version produces output. The autonomous version produces bills.
| Clawbot (autonomous) | Hermes (scheduled) | |
|---|---|---|
| Cost | $80 in one night | $1 per day |
| Output | Nothing | Invoices, briefs, bank reconciliation |
| Guardrails | None | Docker limits, cost ceiling, approval gates |
| Scheduling | Always-on | 9 cron tasks |
| Model | Cheap (DeepSeek-tier) | Cheap (Qwen 3.6 Plus) |
| Reasoning relied on | Heavily | Barely — dynamic tool caller |
| Self-improvement | None | Linear tickets → Claude Code |
Three cost controls matter: run tasks on a schedule to eliminate ambient compute, use the cheapest model that handles your workload (Qwen 3.6 Plus covers 99% of my usage), and add prompt caching. Engineers are already wrecking their 2026 AI budgets by letting agents run unconstrained. The fix is not smarter models. It is boring infrastructure — schedules, ceilings, and a cache.
50 Tickets in 13 Days
Most agent frameworks promise self-improvement through more autonomy — memory, reflection loops, reasoning about past mistakes. Hermes improves through a ticket backlog.
I gave Hermes access to Linear. When a tool returns an error, a cron job fails to deliver, or an API endpoint does not exist, Hermes creates a ticket. Labels it. Sets priority.
In two weeks: 50 tickets created. 39 resolved. Zero open. The breakdown — 40% bug reports Hermes filed against its own MCP tools, 22% feature requests for capabilities that did not exist yet, 12% planning improvements.
AI agents improve through a feedback loop, not through more reasoning. Hermes created 50 Linear tickets in 13 days — bug reports against its own tools, feature requests for missing capabilities. Each ticket gets resolved with Claude Code in the home lab repo. The agent does not fix itself. It documents what broke, and a human fixes the tooling. As of July 2026, 39 of 50 tickets are resolved.
The pattern: Hermes tries to call a tool during a scheduled task. It fails. Hermes files a ticket with the error and context. I open Claude Code, fix the MCP tool, deploy. Hermes works next time.
This is not self-improvement through AI reasoning. It is QA automation with a ticket queue. The agent is its own test suite — and because it runs on cron, it tests itself on a schedule too.
The agent industry sells autonomy. More reasoning, more planning, more intelligence. Build a thing that thinks and it will figure out the rest.
I tried that. It cost $80 and produced nothing.
The agent that saves me hours every month does not think. It calls functions on a schedule. It files tickets when things break. It runs on a Raspberry Pi for $1 a day.
Not exciting. Just useful.
Building AI agents and want to skip the expensive lesson? Get in touch.