Bob Michaels/ai
HALO / Origin9 steps, Nov 2025 to now

How I built the thing that builds everything

HALO is my own method, built and used daily. It did not start as a plan. It came out of a year of shipping real systems and noticing, each time, what went wrong when AI agents did the work.

Why it matters

The volume of what I ship is not luck or long hours. It is what happens when the discipline is a machine.

See the projects page

Dates are first commits read from git. The first step is dated from file timestamps.

  1. 01

    Nov 2025 to Mar 2026

    GEO engines on n8n

    Four generations of a research workflow that asks AI engines buyer questions and scores who gets recommended. All of it ran on hand-wired nodes.

    The lesson

    Research that cannot be repeated cannot be trusted. Citations had to come from real grounding data.

    Carried into 02

  2. 02

    Mar 2026

    Content Lab

    The research became a multi-tenant product. 2,190 commits taught me what happens when many agent sessions touch one large codebase.

    The lesson

    Volume without records is chaos. Every decision needs a place to be written down.

    Carried into 03

  3. 03

    Apr to May 2026

    Grounded Chatbot and Candidate AI Visibility platforms

    Two products started between April 20 and May 27, each with a written charter and a plan and summary per phase.

    The lesson

    A plan a person can reject before work starts is the cheapest quality gate there is.

    Carried into 04

  4. 04

    May 17, 2026

    Manager

    My first orchestrator: several Claude Code agents in parallel, each with a supervising PM agent that approves or blocks its steps.

    The lesson

    Agents need a supervisor that can say no. That idea survived into HALO.

    Carried into 05

  5. 05

    Jul 4 to 5, 2026

    Mission Control

    A second-brain vault and a local dashboard with a terminal per project. The first time status, decisions and handoffs lived in one place.

    The lesson

    If state is not written down at the moment it happens, the next session starts cold.

    Carried into 06

  6. 06

    Jul 23 to 30, 2026

    Skills, vault and consolidation

    A versioned skills library and a doctrine vault, then everything consolidated into one location on July 30. The proposal engine was chartered the same day.

    The lesson

    One place for the method, or it forks and drifts.

    Carried into 07

  7. 07

    Jul 31 to Aug 3, 2026

    Records, learnings and model pinning

    Cycle records, cross-project learnings and warm starts became tools. The model pin turned a policy into a machine that refuses the wrong model.

    The lesson

    A rule written in prose gets skipped. A rule enforced by a tool does not.

    Carried into 08

  8. 08

    Aug 9 to 10, 2026

    Website build engine and the stop gate

    A runbook that only existed as prose became a pipeline. The same week, a stop hook began refusing to let a session end without a written plan and summary.

    The lesson

    The cycle has to be a gate the session cannot walk past.

    Carried into 09

  9. 09

    Sep to Nov 2026

    Adversarial review, page grader, installer

    A second AI engine reproduces defects in the first one's work. A page grader scores every page before it ships. An installer packages HALO for another machine.

    The lesson

    Check the result, not the code.

Where the line ends

HALO, today

HALO runs many client engagements at once with AI agents doing most of the work, and hooks that stop an agent from skipping steps, guessing, or publishing anything unverified. It orchestrates Claude, Codex and a third engine, and one setting changes the model every agent runs on.

commits in the core repository
761
lines of Python in 238 files
80,370
tests, counted
1,564
agent types, one settings file
10

Some gaps are documented in the repo itself: one Codex-side guard stays inactive until a one-time trust keypress, and two live builds still use older agent files.