Expertise
Agent Operations
Running many agents on one codebase without collisions or surprises
Overview
Once more than one agent works on a repository, operations matter: which files each may own, which model tier each task gets, how overnight runs report back, and how a person resumes the next morning with the facts rather than a transcript.
Positioning
Agent operations is the delivery process around the agents. It is where cost, safety, and recoverability are actually decided, and it is mostly plain engineering: manifests, worktrees, hooks, and deterministic reports.
What this means in practice
A phase manifest with file-ownership locks and one worktree per task; cheaper models for workers and stronger ones for checkers; a session-start hook that catches drift; and a morning report generated by a script that calls no model, so the first thing a person reads is true.
Why this matters
An unattended overnight run built a 27,713-hadith collection with three independent checkers and 1,127 passing tests. That works because of the operations around it, not because of any one model.
What we deliver
A work plan that gives each agent its own files
Which model does which kind of task, and what it costs
A morning report written by a script, not a model
Knowledge paths
Related insights
Insight / Agent Operations
An overnight build of 27,713 hadith, and the rule that keeps models out of the rulings
How cheaper agents built Musnad Ahmad unattended in one night, how three independent checkers and a second-vendor plan review caught what they got wrong, and why a scholar's ruling is never filled in by a model.
Insight / Agent Operations
Rebuilding this site with agents, in phases that cannot collide
The repositioning you are reading was carried out by cheaper agents working in parallel worktrees under a file-ownership manifest, with a failing test shown before every change. What the audit found first, how the phases were cut, and what the record shows.