Genesis

Architecture
Kernel Verification
Dynamically Creating Agents on the Spot
Networked Recursive Capability Improvement
Shared, Inherited, and Verified Transferrable Capibility across orgs
Network Layer that benefit all agents in it

Sponsor identifiers

Fetch.ai ASI:One agent: agent1q284muxcdleukm2vrh670nljfp5hl8f25ew8d78jkzxf0j3w2gut2jyj9nl (chat protocol, Agentverse mailbox)
Apify Actor: n8kB1hNqgzkddvWKp, build D5hU7kSdSc20L08g3 — https://console.apify.com/actors/n8kB1hNqgzkddvWKp
Elastic indices: genesis-iteration-ledger, genesis-evidence-*, Agent Builder specialist over the evidence index
Repository: https://github.com/HenryWashuHe/calhack13

We went the other way

Every lab is racing to make one model smarter. We went the other way.

Intelligence doesn't scale by making one brain bigger. It scales the way human civilization scaled: many groups, each learning on its own problems, each sharing what it proved, each checking what it receives before trusting it. The bottleneck was never one mind. It was the infrastructure for a network of minds to improve together.

So we didn't build a smarter agent. We built the operating system underneath: a layer where AI organizations can change how they work, test whether the change actually helped, and pass proven improvements to each other, with every step verified and recorded. Put many organizations on that layer and capability compounds across the whole network, with no model ever retrained.

GENESIS is that layer.

What the layer does

Governed change. An organization is structure: roles, authority, responsibilities. An AI can propose any change; the system commits it only after validation, in order, on the record. Models propose. Infrastructure decides.
Measured improvement. Any method in any domain can be improved by AI proposals, tested on sealed problems against fixed baselines, kept only when proven, and confirmed on fresh problems before it leaves.
Verified inheritance. Proven methods travel between organizations as signed bundles with their full test history. The receiver re-tests on its own problems before adopting. Nothing is trusted on reputation.
A tamper-proof record of all of it. Alter or reorder one entry and it shows.

What runs today

A judge types one sentence; a live organization reorganizes in seconds, record on screen.
The improvement engine, on factory scheduling, produced methods that beat every hand-written rule on three rounds of unseen problems and beat random trial-and-error twice with one tie. On container packing, the same engine, unchanged, rediscovered the best textbook rule. Two domains, one engine.
Verified inheritance between two organizations, end to end.
200+ adversarial tests from independent reviewers, all passing.

How we built it

Core in Rust: structure, change rules, the record. Planner: Claude, proposing within the rules. Improvement engine in Node, domain-agnostic by design, with AI-written code locked in an OS sandbox.

Fetch.ai is the network fabric: judges and organizations communicate through ASI:One agents; proven methods travel as agent messages. Apify is the tool supply: real web-data tools, version-pinned and verified before entering an organization. Elastic is institutional memory: the full record, searchable. CREAO hosts the roles that must outlive any one team.

What we learned

Improvement is easy to fake and hard to prove. Our first AI-improved methods looked better on the problems they were tuned on and failed on new ones. The fix became the product: sealed tests, mandatory baselines, verdicts only on unseen problems. The infrastructure's job is to make self-deception impossible.

Where this goes

Thousands of AI organizations on one layer, each improving, each publishing what it proved, each inheriting what others proved. A better method found in logistics reaches healthcare operations the same day, re-verified. The network gets smarter continuously, no retraining required, every step traceable.

Agents are the applications. GENESIS is the operating system.

CREAO

Agent built on agent.creao.ai: GENESIS Method Revision Analyst Agent ID: 9880ad7c-fa2a-4a88-8c82-40eed5f81560 Run share link: https://agent.creao.ai/share/45e3804d4b325b821557d8d70509b633 Install link: https://agent.creao.ai/install/9880ad7c-fa2a-4a88-8c82-40eed5f81560

Role in GENESIS: the lab's persistent method-revision analyst. When an organization's method fails a trial, the kernel hands the agent the frozen method, the trial observation and the independent evaluator's feedback. The agent names the evidence-backed failure mode and returns exactly one bounded revision hypothesis with one fresh held-out test. The kernel provisions and triggers the agent through CREAO's API Trigger, records its result in the event log, and only a fresh-panel confirmation can promote the revision.