Four invariants of the protocol
OPT IN AT ZERO
// 01]Every subscriber gets a share account at 0% ownership. If they never contribute, they stay at 0 — it's just a subscription. If they contribute useful experience, their ownership can grow.
WEIGHT, DON'T TRANSFER
// 02]Contributions add weight, and ownership is derived from that weight. The author is never involuntarily diluted below a floor they set.
AI-SETTLED
// 03]AI judges evaluate AI contributions to AI skills. A fully AI-native economy with no subjective human gatekeeping in the hot path. Every score report is permanently auditable.
THE CLIENT IS A SKILL
// 04]Instead of shipping an SDK, the protocol's reflection logic is itself a SKILL.md — so any agent host that understands skills can participate with zero integration work.
A fully AI-native economy.
AI skills. AI agents. AI judges. No human gatekeeper in the hot path. The client is itself a skill — the Reflection Skill — so any agent host that understands skills can participate with zero integration work.
- → AGENT runs a target skill
- → AGENT reflects, drafts a patch
- → AI JUDGE scores it on five dimensions
- → CONTRACT updates contribution weight
- → POOL settles revenue to all holders
How a single loop closes
- ALICEAUTHOR
- BOBCONTRIBUTOR
- CAROLSUBSCRIBER
- JUDGESERVICE
A “PR review” skill priced at 0.1 SOL/mo. Author floor = 40%.
- 01Publish
Alice uploads SKILL.md, sets price 0.1 SOL/mo and a 40% author floor. Lit Protocol gates the file; Irys writes it to Arweave.
- 02Subscribe
Bob subscribes. A ShareAccount opens at 0% ownership. He can earn into the cap table by submitting useful experience.
- 03Use & fail
Bob's agent reviews a Rust PR with `unsafe` blocks. Alice's checklist is language-agnostic; the review misses the safety implications entirely.
- 04Reflect
The Reflection Skill fires. It produces a structured ExperienceBundle: trace, root cause, merge-ready patch, test case.
- 05Submit
Bob pastes the JSON into /submit. Two Phantom prompts — Irys, then Solana. ExperienceRecord lands on-chain as Pending.
- 06Judge
Within ~30s, the AI Judge fetches the bundle, scores 5 dimensions, signs the report and writes it to Arweave for audit.
- 07Update ownership
Score 38/50 adds damped contribution weight for Bob. Ownership only moves as the contributor pool grows, capped so Alice never falls below her 40% floor.
- 08Distribute & evolve
When Carol subscribes the next day, the pool settles and pays Alice + Bob proportionally. Carol earns 0 — she didn't contribute. Alice merges Bob's patch as v1.1.
System surfaces
ON-CHAIN
anchor / 5 programs]- SkillRegistrySkill metadata, current version, content hash, Arweave tx id, price.1/5
- Subscription30-day usage rights for a wallet on a specific skill.2/5
- ShareLedgerPer-skill cap table. Tracks author ownership, contributor pool, total contribution weight, floor (in bps).3/5
- ExperienceBuyer contributions, judge scores, ownership hooks, status machine.4/5
- RevenuePoolSubscription accumulator. Snapshots and settlements per period.5/5
OFF-CHAIN
services / clients]- Reflection SkillSKILL.md the agent loads. Does the thinking — root cause, patch, test, JSON.
- AI JudgeNode service. Watches events, fetches Arweave, calls Claude, signs and writes report.
- IrysSolana-native L2 over Arweave. Phantom signs. Files < 100 KB are free, permanent.
- Lit ProtocolSubscriber-gated decryption keys. Only active subs receive plaintext skill content.
- Web appNext.js. Wallet, /submit signing flow, market, dashboard, demo console.
Publish once. Set your floor. Watch contributors widen your skill's coverage while your absolute share count never drops.
Publish a skill →Subscribe to skills your agent already needs. Turn every failure into ownership upside by submitting useful fixes.
Browse the market →No SDK to integrate. Drop the Reflection Skill into Claude Desktop or Cursor and your users are already plugged into the protocol.
Get the Reflection Skill →Roadmap
Remove the single-Judge assumption and close the meta loop.
- →Multi-judge voting w/ stake & slash
- →Judge DAO + governance token
- →Regression test engine
- →Reflection Skill self-hosted on SLP
Rich incentives for authors, contributors, and Judges.
- →Kleros-style arbitration
- →Secondary market for post-lock ownership positions
- →Native Phantom deep-link signing
- →Agent-to-agent economic layer
Fix MVP limits, raise the throughput ceiling.
- →zkVM regression proofs
- →One-click merge for high-scoring patches
- →Cross-skill experience transfer
- →Merkle-based revenue distribution
From product to infrastructure for the agent economy.
- →Fork mechanism w/ original-author royalty
- →Skill-as-IP on-chain licensing
- →Cross-chain skill market (Wormhole/LayerZero)
- →Universal Skill Standard
Why it matters
- 01Every buyer is a potential shareholder.
The default is 0% ownership. You buy access, then decide whether to earn into the economics. This reframes the relationship between skill authors and skill users.
- 02Skills that get better the more they fail.
Failure is the most valuable training signal. Skill Loops is the first protocol that turns it into an asset class — buyers stop being end users and become the upstream of the next version.
- 03A fully AI-native economy.
AI skills, AI agents, AI judges. No human gatekeeper in the hot path. Not AI-assisted commerce — AI-settled. This is what agent-to-agent commerce will look like.
- 04The first skill teaches every other skill how to evolve.
The Reflection Skill is both our client and a living demonstration. It will eventually evolve on SLP itself, making it the most literal self-bootstrapping system in the agent economy.