HALO
Human · Agentic · Lifecycle · Orchestration
The multi-agent system I run client work on
I direct the work. AI agents do most of the typing. HALO stops them from guessing or shipping anything unchecked. It's how I run a team of AI agents to deliver enterprise systems.
Measured from the repository
761
commits in the core repository
80,370
lines of Python in 238 files
1,564
tests, counted, not run
10
agent types, one settings file
I shipped 62 builds and tools in the last 24 months. HALO is why I ship this fast. AI agents wrote much of the code. My part is the specification, the direction, the review and the shipping. Counts are measured from disk.
AI governance
What HALO does
I run many client engagements at once. I can't hold a dozen projects in my head, and AI agents left alone guess and publish things nobody checked.
HALO fixes both. One cockpit shows every project and whether it's working, idle, or waiting on me. Every job gets a written plan before it starts and a written summary when it ends. A second AI engine tries to break the work before it ships.
It runs Claude, Codex and a third engine. One setting changes the model every agent runs on, and the tools refuse work on the wrong one.
Every mistake gets written down as a lesson, and every agent reads the lessons before it starts. More than one hundred so far. Each mistake is paid for once.
Workstreams
4 liveApprovals
7
awaiting your word
Deploys
verified · reversible
Agentic workflows
How a job runs with a human in the loop
Plan
Every job starts with a written plan. I can reject it before any work starts.
Build
Agents do the typing. Hooks stop them from skipping a step or publishing anything unverified.
Break
A second AI engine gets the uncommitted work. Its one job is to reproduce defects.
Ship
A session can't finish without a written summary. I review the result and ship it.
I approve every stage, and nothing ships without my word.
Read more