Agents Collab
An experimental skill for configured model collaboration and independent review,
plus the lab notebook of how it came about.
Status: experimental, pre-release. The skill comes from a Codex-led pilot.
Licensed under the MIT License.
Source ·
Website and experiment journal.
No versioned release has been made.
Product target and current limitation. The goal is that a lead agent uses
Agents Collab for appropriate configured teamwork without the user naming the
skill. That implicit activation is not achieved: in the
public routing gate, ordinary
prompts missed both required collaboration routes. Explicit invocation
(Use $agents-collab ...) matched all three expected routes in that small
diagnostic and remains the supported fallback, not the target. A
draft activation study defines the
proposed next measurement. Method review passed for the earlier draft; the
18-run study remains unexecuted, unapproved and unpreregistered. The source
now carries the experimental activation description, which does not establish
dependable implicit activation.
Quick start
Install for Codex with Vercel’s Skills CLI:
npx skills add jaredchu/agents-collab --skill agents-collab --global --agent codex
Installation requires Node.js 22.20+ (with npx) and Git; profile setup and
validation also need Python 3.9+. Running the command again can replace existing skill
files; use the manual route to preserve them.
--global installs to your user skills directory (~/.agents/skills), outside
project repositories. It does not create or overwrite your collaboration profile.
After installing or updating the skill, create your profile with the wizard, which asks for your writer,
reviewers and reasoning requests, previews the result and writes only after you
confirm:
python3 ~/.agents/skills/agents-collab/scripts/setup_profile.py
It keeps an existing profile unless you pass --replace; --dry-run previews
without writing. It asks for no credentials, calls no model and cannot check
that a model or reasoning level is available to you. You can also
copy a template and edit it by hand
and validate its structure. Invoke it with:
Use $agents-collab to implement this change with independent review.
Installation alone does not configure model access or enable automatic routing.
A public routing gate missed both
required collaboration routes on ordinary prompts. Explicit skill invocation
matched all three expected routes in that small diagnostic, including staying
solo for routine work; no model names were supplied in the prompts.
A completed six-task quality pilot
found a reproduced regression corrected by both independent-review paths and
missed by self-check. No quality advantage over matched manual coordination was
established for the skill. Review-versus-self-check thresholds are sensitive to
an existing test failure supplying a second initially material draft; the report
preserves the primary counts, sensitivity and recovered grading receipts.
Prefer no Node.js? Use the manual ZIP or checkout instructions.
The workflow currently targets a Codex lead; installing the files for another
agent does not establish support for that lead.
What it is
skills/agents-collab/ is an agent skill: Markdown instructions, a reference
contract, inert templates, configuration helpers, an optional CLI adapter and
installer, and a copy of the MIT LICENSE. It
helps a
lead assistant session:
- choose a writer and independent reviewers from a personal profile kept
outside every project (
~/.config/agents-collab/profile.json);
- send reviewers a neutral packet and the same frozen artifact, not the writer's
conversation or each other's first findings;
- cap review rounds, record dispositions, and keep execution records in a
private records directory rather than in project documentation;
- hand only durable outcomes (decisions, verified behavior, risks, remaining
work) back to the project.
The optional Rust CLI backend records local handoffs and immutable
packets. Install its separately packaged executable and explicitly enable it;
installing the skill alone keeps the manual workflow. Backend preferences live
in ~/.config/agents-collab/backend.json, leaving profile.json unchanged.
The CLI dispatches no models. Explicit owner-approved quota continuations
retain the same unfinished stage and history without changing profile defaults.
Support is currently limited to macOS arm64 and
local APFS; see installation and use.
What it is not
Read these before relying on it:
This is the beginning of the README. Read the rest at the original. Read the full README ↗
