PhantomByte Publishes Agent Automation Toolkit: 92 Scripts Behind a Two-Agent AI System

TelAve News/10911385
The MIT-licensed repository shows how two AI agents plan, run, and verify their own scheduled work, with every script written to catch a run that reported success it did not have.

RONKONKOMA, N.Y. - TelAve -- PhantomByte has published Agent Automation Toolkit, a public repository containing the scripts, scheduled jobs and skill definitions that run two AI coding agents as working automation rather than a demo. One agent runs on Windows and one runs on Linux, and both keep to the same rule: verify before you report.

The repository holds 92 runnable scripts grouped by domain, 34 scheduled jobs with their expressions and full instructions, and 281 skill definitions, 262 for Hermes and 19 for Pi, plus an architecture document on how the two agents hand work to each other. Each job declares a runner type, and 12 of the 34 run a deterministic script with no model call at all, which keeps the cheap half of every schedule cheap. Three hooks run before a tool executes: a per-turn tool-call ceiling, a guard that moves a stale scratch file aside so a legitimate write is not refused, and a stripper for hard-wrapped delivered text.

More on TelAve News
Most of the code exists because a previous run claimed a result it had not produced. An email job once reported success with the attachment missing, so imap_readback.py now re-reads the inbox and compares the decoded attachment length to the file on disk. A nightly backup failed for consecutive runs behind a pre-push guard, so the backup script gives every git call a timeout and a non-zero exit, which makes a failed push a failed run. The job-search pipeline resolves each posting to the employer's own apply page and fetches it live, because a blocked fetch is not a verified lead.

"Every script in that repository exists because a run told me it worked and it hadn't," said Vinny Barreca, systems builder at PhantomByte. "A status line that prints every time stops being read, so most of these jobs stay silent until something is actually wrong. The ones that must speak are written to say so."

More on TelAve News
The toolkit is aimed at developers running local or self-hosted agents, at teams moving work out of a chat window and into a schedule, and at engineers who inherited automation nobody documented. It runs on Python 3.9 and up with the standard library for most of the code, adds playwright, requests, scrapling and weasyprint only where a script calls for them, and needs Node 18 for two files.

Repository: https://github.com/vincentsativa/agent-automation-toolkit

About PhantomByte: PhantomByte is an independent engineering studio in Ronkonkoma, New York, building production AI systems, sovereign AI infrastructure and hands-on technical tutorials.

Articles: https://articles.phantom-byte.com

Source: PhantomByte
Show All News | Disclaimer | Report Violation

0 Comments

Latest on TelAve News