30 Days of a Local AI Agent on My Own Machine: What Actually Broke

Description and use

Automating tasks with a local AI agent failed due to environmental issues. The solution involved idempotent steps, logging, and human approval gates.

Problem

An AI agent failed to reliably automate tasks on a local machine due to environmental issues rather than model limitations.

Outcome

Reliability improved by focusing on environmental stability, idempotency, logging, and human approval gates rather than model upgrades.

Workflow

• Run agent locally
• Agent drives screen, shell, browser, and Android via ADB
• Agent performs 400 tasks
• Identify failures: stale UI, leaks, prompts, partial writes
• Add idempotent, logged steps
• Add human approval gate for irreversible actions
• Resume tasks instead of restarting

Evidence

Every serious failure I hit was environmental: a stale browser session, a lock file from a crashed run, a port already in use, a half-typed shell command left by the previous task.

Source

The single biggest improvement was not a better model. It was making every step idempotent and logged, so a task could be resumed instead of restarted.

Source