THE GAIA ECOSYSTEM

One system.
Four names.

Gaia is an open capability control system for AI agents. It is built as four named parts, and this page has exactly one job: tell you what each name is and how they relate. Every part keeps its own documentation — this page links to it and never repeats it.

What is Gaia?

Agents can now do far more than any one person can verify. Gaia’s answer is to stop treating that as a single problem and split it into four questions that are usually mashed together:

  1. What capabilities exist, and why should I trust them?
  2. What do they actually cost, and do they work?
  3. What should enter this session, right now?
  4. How far can that be pushed before it stops being safe?

Each question has a name. Keeping them apart is the whole design: one of them is a permanent public record, one is a laboratory, one runs inside your session, and one is not open yet.

THE FOUR NAMES

Who does what.

THE RECORDACT ACTIVE

Skill Tree

The permanent public record of what agent capabilities exist, who demonstrated each one first, and what evidence stands behind it. A skill is not an entry in a list — it is a record with a trail: its origin contributor, its evidence, its stars, and a timeline of how it earned them.

You want this if you are looking for a capability, checking whether one is trustworthy, or putting your own work on the record.

Browse the Skill Tree
THE LABORATORYACT ACTIVE

Research

The open laboratory that measures what capabilities really cost and whether they help — what a skill costs while merely listed, what it costs when invoked, and how different harnesses behave — then publishes the method along with the findings that went the other way. You are reading it right now.

You want this if you want the numbers behind a claim, or the method to reproduce it yourself.

THE RUNTIMEWIP EXPERIMENTAL

Skill Heaven

The layer that decides what enters an agent’s session and how much of it. Where the Tree answers what exists and why trust it, Skill Heaven answers what should be in this context, right now — one ladder from off to max, with a stated posture at every rung and the dose measured rather than guessed.

You want this if your agent’s context is crowded with capabilities it is not using on this task, and you want that priced instead of assumed.

Experimental. The measurement is public before the product is. There is nothing to install from this page yet — when there is, it will say so here.

Read the benchmark method
A GATED TIER OF SKILL HEAVEN🔥 HELL · GATED

Skill Hell

The other direction on the same ladder. Where Heaven admits a quiet session, Hell summons everything the evidenced world has to offer for long autonomous runs — still bounded by the rung you chose. It is a tier of Skill Heaven, not a separate part of Gaia.

Gated, and staying gated until measured trust coverage clears the gate. There is nothing to install, nothing to join, and no waiting list. When it opens, this is where it will say so.

HOW THEY RELATE

It is a loop, not a stack.

  1. Research measures cost, quality, failure, and harness behaviour.
  2. The Skill Tree records which capabilities are trusted, on what evidence, and by whom.
  3. Skill Heaven uses that record to decide what enters a session, and at what dose.
  4. The run itself produces new evidence — what it cost, what it fixed, what it broke — which goes back to Research.

Each rotation leaves the next one better informed. That is why the names stay separate: measuring, remembering, and admitting are different jobs. Fuse them and the thing making runtime decisions ends up grading its own homework.

START HERE

Which one do I want right now?

Reader intent mapped to a Gaia surface, its true state, and one concrete action
If you want to…Go toStateDo this
Find a capability, or check if one is trustworthySkill TreeACT ACTIVEBrowse the Skill Tree ↗
Put your own skill on the public recordSkill TreeACT ACTIVERead the contributor path ↗
See the numbers behind a claim, or reproduce oneResearchACT ACTIVEOpen the research ledger →
Understand what a crowded context costs youSkill HeavenWIP EXPERIMENTALRead the benchmark method →
Run fully autonomous, everything admittedSkill Hell🔥 HELL · GATEDNothing yet — it is gated on evidence, not on demand.

WHAT CHANGED

If you learned Gaia earlier.

REVISION 1 · 2026-07-29

Some of what Gaia shipped this year replaces things people already learned. This is a standing surface, not an announcement: when something makes prior knowledge wrong, it gets a dated entry here.

The four names 2026-07-28

Gaia is now told publicly as four names — the Skill Tree, Research, Skill Heaven, and Skill Hell. Nothing you already use changed; what changed is the map. Skill Heaven is named as the runtime layer and is experimental. Skill Hell is named and gated.

Yggdrasil II 2026-07-07

The Skill Tree’s structure changed. If you learned it before this date, three things you know are now wrong: skill types collapsed from four to two, the progression branch is derived at read time and is never written down, and the rank words above three stars were renamed. Trust Magnitude became the single gate for ranking up.

Every claim has a trail.

This page owns the relationships and nothing else. Each part of Gaia keeps its own thesis, its own docs, and its own receipts — and this is the page that tells you which one you are standing in.