If Everything Can Be Recreated, Who Gets Remembered?
When AI agents make software implementation reproducible on demand, original authorship does not disappear. Why Gaia Skill Tree and Yggdrasil preserve verifiable lineage and evidence-backed trust.
PUBLIC RESEARCH SIGNAL
Pushing the limits of agent capability.
Gaia Research is the open laboratory for evidence-first agent work. We observe, benchmark, verify, and publish the frontier — every claim links back to its receipts.
MILIM · CHIEF CAPABILITY SCOUT
VERIFIED EVIDENCETWO LABS, ZERO SIGN-UPSBUILDING IN PUBLIC
FROM THE BLOG
A little field research.
Short, human-readable notes from the Gaia lab—where the work gets a little strange before it gets useful.
Capability Amplification: When Better Scaffolding Beats a Bigger Model
Structured agent scaffolding can buy task-level reliability more cheaply than buying a larger model. Why procedural skills are an economic instrument rather than prompt debt.
The Return of the Simple Reflex Agent: Why Jev Is a Classifier, and Why That Matters
Jev makes fast typed decisions instead of generating prose. Its useful return to reflex agents and probabilistic classification puts frontier LLM calls and local classifiers on the same audit list.
PULL A LEVER, SEE WHAT SPARKS
Labs & Games.
Little browser toys we build to poke at agent capability — no login, no download, nothing leaves your machine. Go ahead, boss. Break something.
Live · Game
Infinite Skill Craft
Smash two dev skills together and see what the forge coughs up. Then do it again. And again. Infinite, obviously.
Start craftingBeta · Tool
Context Diet
Feed it a bloated context file and watch it shrink to a lean, mean skill sliver. Weigh-in stays in your browser.
Trim a promptHELL HEAVEN INDEX · ARC I VERIFIED
Stop installing skills.
Start summoning them.
Marketplaces make you install skills forever — bloat you never asked for, pinned to every repo. Skill Zero is the launcher that starts a harness with zero ambient skill debt. Heaven and Hell are the summon directions on the HH axis: Heaven converges toward the right few skills; Hell explores the evidenced world.
New to Gaia? Start with the three names — Gaia Skill Tree, Gaia Research, and Gaia Skill Heaven (housing Skill Zero and Skill Hell). The HH Index is in the works.
Converge on the right few.
Heaven is the curated summon direction: it pulls the evidenced pool toward the small set a design session actually needs. Skill Zero gives it a clean launch when useful; Heaven decides what deserves to come back.
Full gas, autopilot.
Summon every good skill in the evidenced world for autonomous fleets and long loops — bounded by the rung you picked. Unlocks only when the registry’s trust-coverage clears a measured gate. Ludicrous mode ships with a seatbelt.
Skill Zero is the usable launcher prototype. The Hell Heaven (HH) Index — a per-skill hellHeaven stamp over Heaven/Hell polarity — is the research now in the works (routing currently falls back to relevance ranking). Read the benchmark method & receipts → · Vision ↗ · Mission ↗.
Skills you can install today.
Local-first Claude Code / Cursor / Windsurf skills. One line to install with npx skills — they run against your files and never upload their contents.
context-diet
WIP EXPERIMENTALMeasure and compact an oversized agent-context file under the harness limit without dropping a single rule.
npx skills install gaia-research/skill-context-dietci-churn
ACT ACTIVEGitHub Actions cost analyzer and flaky-test detector: measure wasted CI compute and prevent retry-push churn.
npx skills install gaia-research/skill-ci-churncost
ACT ACTIVEMulti-harness token-usage cost reporter for pi, Claude Code, Codex, and opencode session logs.
npx skills install gaia-research/skill-costfuse
ACT ACTIVECompose two installed agent skills into one unified SKILL.md — the composition engine behind the Skill Tree.
npx skills install gaia-research/skill-fusescout-fleet
ACT ACTIVEParallel cheap-scout fan-out harness (2x, 4x, or cascaded funnel) with zero-token RRF deduplication.
npx skills install gaia-research/skill-scout-fleetimage-tuner
ACT ACTIVEInteractive bounding-box image positioning, zoom, and framing tuner with live coordinate and CSS export.
npx skills install gaia-research/skill-image-tunerClaims deserve a trail.
| Research item | Type | Status | Evidence note |
|---|---|---|---|
| Orchestration Guide | GUIDE | WIP EXPERIMENTAL | Step-by-step, plain-English orchestration playbook: when not to orchestrate, the one-fast-lane rule, receipts over raw output, and keeping the planner's cache warm. Experience-based — no control study yet. |
| Context Diet | LAB | WIP EXPERIMENTAL | Local token-budget estimator; comparative benchmark results pending review. |
| The Compounding Cost of CI Failures | POSTMORTEM | VRF VERIFIED | Postmortem of Epic #780 introducing CI Churn as a first-class cost metric. |
| Agent Cost Reporting | RESEARCH PLAN | PRP PROPOSED | Proposed study of the gap between agent estimates, rate-card totals, and invoices. |
| Gaia MCP | PRODUCT | ACT ACTIVE | Model Context Protocol server exposing the Gaia Skill Tree to Claude Code, Codex, and Cursor. |
| Scout Fan-Out | BENCHMARK | VRF VERIFIED | Empirical study of parallel cheap-scout fan-out vs single mid-tier scouts across 360 runs: Pareto frontier, prompt-cache amplification, and flake rate reduction. |
| HH Benchmark | BENCHMARK | VRF VERIFIED | Drug-trial method for scoring skills by marginal efficacy against model baselines. Arc I verified: baseline floor pricing, census, and machine-gated claim index. |
| Context Compaction Bench | BENCHMARK | VRF VERIFIED | Empirical study across 37 live agent sessions (25 across S1/S2/S4/S6 + 12 across S3) and 8 autocompaction thresholds on Gemini 3.8 Flash: Pareto frontier, prompt-cache amplification, reacquisition thrashing, and super-linear reasoning scaling. |
RESEARCH SIGNAL / PERMANENT RECORD
Let the research take root.
Gaia Research tests the frontier in public. Gaia Skill Tree keeps the attributable, evidence-backed record of what agents can actually do.
- FROM
- Open research, methods, and experiments.
- TO
- A public registry of demonstrated agent skills.
A field note from Milim, Chief Capability Scout.
Make the science
louder than the hype.
- 01 Share the method. Let others inspect the work.
- 02 Verify before you flex. Evidence beats vibes.
- 03 Build safe, robust agents. Then publish the receipts.