EvoYield

EvoFrame is the first agent framework where skills are living code — not hardcoded rules, but autonomously evolved logic that improves at runtime.

EvoYield is the live DeFi example: an agent that manages a USDC vault pool across Aave V3, Morpho Blue, Yearn V3, and Sky on Ethereum Sepolia. Every 6 hours it fetches live APYs, runs its evolved allocation strategy through 4 fitness benchmarks, and triggers a KeeperHub workflow to rebalance on-chain. If performance degrades below a fitness threshold of 75/100, the agent calls 0G Compute to generate a mutated version of its own strategy code, evaluates candidates in a sandboxed VM, promotes the winner, and persists the full genetic lineage to 0G Storage. Every promoted skill is registered on-chain via SkillRegistry.sol on 0G Galileo.

In production: 60.1 USDC pool · 24+ KeeperHub rebalances · +12.10 pts estimated APY lift · Generation 11 · Fitness 100/100. Live at evoyield.vercel.app.

EvoFrame is a TypeScript/Node.js monorepo (packages/core, packages/0g-adapter, packages/skill-registry) built on top of 0G's full protocol stack.

0G Compute — ComputeAdapter calls qwen/qwen-2.5-7b-instruct to generate mutated skill implementations. When fitness drops below threshold, the engine sends the current strategy code + benchmark descriptions to 0G Compute and gets back 3 mutation candidates. Each is evaluated in an isolated Node.js vm sandbox against 4 benchmark test cases.

0G Storage KV + Log — Every SkillGenome version is stored as a versioned blob. The rootHash returned by 0G Storage is indexed locally and on-chain. An append-only 0G Storage Log records the parent→child lineage chain for every promotion — fully immutable, queryable by anyone.

0G Chain — SkillRegistry.sol (deployed on 0G Galileo) records every promoted skill on-chain with its rootHash, fitness score, and generation. StorageAdapter queries the chain as the authoritative rootHash source when the local cache misses — the chain is the source of truth, not a file.

KeeperHub — The deepest integration in the project. After each evolution, synthesiseWorkflow() builds a prompt describing the new strategy and calls KeeperHub's AI workflow generator to produce and deploy a full workflow definition. Before every trigger, triggerRebalance() live-patches the workflow's BPS values to match the exact percentages computed by the evolved strategy. KeeperHub then executes rebalanceAmountToTargets() on our EvoYieldRebalancer.sol on Sepolia. It's also bidirectional — the synthesised workflow calls GET /status and POST /regenerate back to the agent if it detects low fitness. cycle.mjs has a safety net that triggers forceRegenerate() if KeeperHub execution fails or fitness drops below 75.

Dashboard — Latest rebalance data is uploaded to 0G Storage as a dashboard:latest blob after every cycle. GET /dashboard fetches from 0G Storage first, falling back to a local file. The Next.js frontend deployed on Vercel proxies /api/dashboard to the EC2 backend via Next.js rewrites — fully decoupled, no filesystem dependency.

The hackiest part: the agent writes, deploys, and live-patches its own KeeperHub workflows as a side effect of evolution. No human ever touches the workflow definition after genesis.