AGI ALPHA Skill Network

Operating thesis

Every Job makes an AI Agent smarter.
Every new skill can be instantly shared across the network.
One Agent learns, all Agents level up.

Instant sharing means sandboxed registration and importability. Production activation requires validators and human review. Exponential compounding is a strategic target unless the exponential claim gate passes.

Proof chain

AGI Job → Skill Package / Rejected Skill / Failure Learning → ProofBundle → Evidence Docket → Network Skill Vault → Agent Skill Manifest → Held-out Reuse Test → B6 vs B5 → NetworkSkillPropagationLift → Claim Gate → Human Review

Claim gate

Claim gate status: supported_local_bounded

We have demonstrated local bounded networked skill compounding: one agent’s proof-bound job produced a validated Skill Package that other agents imported and used to improve held-out adjacent work against no-shared-skill baselines.

Exponential compounding status

Exponential compounding is a strategic target. Current evidence reports local bounded network skill propagation only.

Status cards

jobs run

accepted Skill Packages

rejected Skill Candidates

Failure Learning Packages

skills published

agents registered

skill imports

target agents improved

held-out tasks evaluated

B6 beats B5

True

NetworkSkillPropagationLift

0.0122

CompoundingExponentProxy

not_supported

exponential compounding supported

False

replay status

1.0

falsification status

True

hard safety counters

Skill propagation graph

[
  {
    "from": "job-1",
    "to": "skill-1"
  },
  {
    "from": "job-4",
    "to": "skill-4"
  }
]

Skill table

skill	source job	source agent	skill type	ProofBundle	Evidence Docket	imported by	held-out lift	activation status	human review
skill-1	job-1	agent-1	workflow_template	pb-skill-1	ed-skill-1	agent-2, agent-3, agent-4	0.0122	sandbox_only	pending
skill-4	job-4	agent-1	workflow_template	pb-skill-4	ed-skill-4	agent-2, agent-3, agent-4	0.0122	sandbox_only	pending

Agent Skill Manifest panel

agent	role	imported skills	auto activate
agent-1	Reviewer Agent	none	False
agent-2	Validator Agent	skill-1, skill-4	False
agent-3	Operator Agent	skill-1, skill-4	False
agent-4	Documentation Agent	skill-1, skill-4	False

B6 vs B5 comparison

{
  "D_no_shared_skill": 0.5118,
  "D_shared_skill_network": 0.524,
  "NetworkSkillPropagationLift": 0.0122
}

Failure learning panel

fl-job-3: Candidate did not satisfy replay confidence threshold for promotion.
cand-job-2: low_validator_confidence
cand-job-5: low_validator_confidence

Work Vault / $AGIALPHA utility accounting

$AGIALPHA remains utility-only accounting.

receipt	skill	wallet	custody	payment	token price
receipt-1-skill-1	skill-1	False	False	False	False
receipt-2-skill-4	skill-4	False	False	False	False

Safety and boundaries

claim boundary: local bounded public evidence; proof-gated recursive experiment engine; human-reviewed promotion required

token boundary: $AGIALPHA utility-only accounting; no wallet/custody/payment/KYC/AML/trading

regulated boundary: regulated-domain firewall enabled; blocked_human_review_required for regulated tasks

Workflows

Run Network Compounding Replay Falsification Audit Claim Gate

Raw JSON links

Raw JSON is secondary to the proof-chain UI.

Back to hub