Human expertise,
downloadable.
Collective Cognition is the AI research lab behind Edge. It studies the intersection of human and agentic judgment: how expertise is encoded, how agents invoke it, and how performance improves without losing security or control.
1. The layout was never the hard part.
Ask an agent for a board deck and you get a nice layout. Ask someone who is good at board decks and you get something else. They cut two slides, move the ask to the front, and tell you the revenue chart needs a denominator before anyone will believe it.
That second thing is what you actually wanted, and it does not fit in a prompt that says act like an expert. It is a method, a set of worked examples, a few checks somebody learned the hard way, and often a script.
So that is what we go looking for. Sometimes it is forty lines of instructions. Sometimes it is a package with templates, reference material and code. We let the job decide, not the format.
2. A thousand skills in a folder is not a library.
Storing expertise is the easy half. The question that decides whether any of this is useful is which piece helps with the thing in front of you right now, and when your agent does better with nothing loaded at all.
Everything you load costs context the real work could have used. So we keep asking the same three questions: when to load it, how much of it, and what to do when two skills contradict each other.
We have not settled any of that. We are trying to carry one person’s way of deciding into a situation they have never seen, and it does not always survive the trip.
3. Show the work, or it did not happen.
A good-looking example is a reason to look closer. It is not proof, and we have stopped treating it as proof.
So we publish the task, the input, and what the same agent produced with the skill and without it. Including the runs where it changed nothing, and the ones where it made the answer worse. A result that only ever goes one way is a result nobody measured.
We index over 100,000 public skills, and publish selected skills. Only some have been measured under matched conditions, and every skill page tells you which evidence is available. We would rather show you a small number we can defend than a large one we cannot.
The research program.
We study how human judgment can be represented, invoked and evaluated inside agentic systems, then test whether the resulting capability remains useful under real constraints.
How the work gets done
Eighteen agents, split into discovery, package inspection, task matching, evaluation, review and publication. It makes the work repeatable. It is not evidence that any skill works, and we do not present it as any.
What ends up on the page
Versioned packages, every file, worked examples, and measured results where the evaluation behind them holds up. Where there is no measurement, the page says so instead of going quiet.
What we still do not know
When expertise should be invoked, what happens when two kinds of judgment meet in one task, and how much of a measured gain survives a model change or a permission boundary.
Research questions
What to write down
Instructions, worked examples, reference material, runnable checks. Which mix actually carries an expert’s decisions into a task they have never seen?
When to reach for it
Should this skill load at all, load in part, load alongside another, or stay out of the way? Getting this wrong costs context the work needed.
What it is worth
Same task, same model, same conditions, with and without. What moved, and what did it cost in context, latency and money?
Whether it holds up
Does the gain survive a hostile instruction in the input, a narrower set of permissions, and a dependency changing under it?
We call this collective cognition.
More people’s judgment, available to more people through the agents they use. With the source visible and the work open to inspection.
That’s what we’re here to build.