Notebook · November 2026 · Jared Chu

Personal collaboration habits don’t belong in project files.

Agents Collab is an experimental agent skill for choosing a writer and independent reviewers from a personal profile, and for keeping execution records private. This site is the notebook of how it came about: an earlier setup that drifted, research showing the “agent team” idea already existed, two small trials that missed their targets, a public routing gate, a recovered quality pilot, and a cleanup that showed why the boundary matters.

Release status

Pre-release. The skill comes from a Codex-led pilot. It is licensed under the MIT License, and its source is at github.com/jaredchu/agents-collab. The guide and notes below are complete and readable now.

Install and configure it Read the research notes

Install for Codex

shell · install for Codex

npx skills add jaredchu/agents-collab --skill agents-collab --global --agent codex

Installation needs Node.js 22.20+ and Git; profile setup and validation also need Python 3.9+. Running it again can replace existing skill files; the manual route preserves them. The skill installs in your user directory; your personal profile stays separate. Manual installation without Node.js is also available.

Then create your profile interactively. The wizard asks for your writer, reviewers and reasoning requests, previews the result and writes only after you confirm. It keeps an existing profile, asks for no credentials and calls no model:

shell · create your profile

python3 ~/.agents/skills/agents-collab/scripts/setup_profile.py

It cannot check that a model or reasoning level is available to you. Prefer editing JSON? Copy a template by hand. Then invoke $agents-collab.

The idea in one pictureTwo kinds of knowledge

A project needs to know what was decided and why. It does not need to know which model one person prefers to write with, at which reasoning level, or where that person keeps review receipts.

Shared project knowledge

in the repository · read by everyone

Decisions and the reasons for them
Verified behavior and checks that ran
Material risks and open questions
Genuine project requirements, such as a required reviewer
Attribution where the project needs it

Private configuration and records

~/.config/agents-collab/ · records_root

Preferred models, transports and reasoning requests
Routing, review recipes, round caps, time targets
Dispatch receipts and individual review findings
Resume state for an interrupted collaboration
No credentials, ever

The skill reads the private side to decide who writes and who reviews, then hands only durable outcomes across the boundary.

What it isA procedure, not a runtime

The package is a small folder: instructions, a configuration contract, a review protocol, two inert templates, an example profile, interface metadata, two dependency-free configuration helpers (a profile validator and a setup wizard) and the MIT LICENSE. A lead assistant session follows the instructions with its own tools; neither helper calls a model.

Prompt-led. No automatic dispatcher, snapshot freezing, hard-deadline implementation or enforcement layer.
Codex-led pilot. Other lead assistants and adapters are not validated.
Accepted is not attested. An interface accepting a model or effort request is not runtime proof of it.
Fresh is not isolated. A fresh reviewer context is not filesystem isolation.
Structure is not access. The validator never checks that a model exists or that you can use it.

So farWhat the trials showed

Coordination threshold: not met, twice

A completed six-task quality pilot found a regression corrected by independent reviews and missed by self-check. No added quality advantage over matched manual coordination was established for the skill. The review-versus-self-check result is sensitive to existing test failures and the small number of initially material drafts.

The public routing gate matched 1/3 expected routes on ordinary prompts. Explicit skill invocation matched all three in a separate diagnostic, without naming models. The planned 36-run comparison stopped; reliable implicit activation remains unestablished.

Two earlier five-task pilots compared an existing workflow with and without the skill. Neither met the 20% coordinator-improvement rule fixed locally before each trial. Some writer-time and quality figures moved, in different directions and with confounds. The defensible conclusion is narrow: separating personal settings from project files is workable. It is not a productivity result.

Read the experiment notes and their limits.

NotebookEntries

Mid-October → November 2026Story: from copied model rules to a profile
The earlier per-project setup, the drift it caused, and the cleanup afterwards.
7 November 2026Research: the team that already existed
Why a fixed architect/builder/reviewer team was dropped for a narrower problem.
9 November 2026Experiments: recovered quality pilot, routing gate and earlier trials
Frozen public inputs, completed grading, primary counts and sensitivity, and earlier negative results.
ReferenceGuide: install, configure, invoke
Copy commands that do not overwrite anything, the profile setup wizard, a template and the validator.