Introduction

An AXL-native multi-agent debate for onchain LP decisions.

Zuno is an AXL-native agent mesh for onchain operations. The first workflow is Uniswap v4 liquidity management, but the product is the mesh: peers, transport, and adversarial reasoning before reviewed execution. It is built around a Turnkey-backed Zuno wallet - the user’s main wallet stays separate, and the agent wallet is the execution wallet.

It runs in your terminal. Two flows live behind the same debate:

• Create a brand-new position - point Zuno at a token + amount, the Scout surveys live pools on the chain, the Strategist proposes 2-5 pool/range candidates, the Critic stress-tests them, the Arbiter resolves on deadlock. See CLI / create.
• Rebalance an existing position - the same four agents debate whether and how to move its range. See CLI / plan.

Both flows share Scout, Strategist, Critic, Arbiter and converge on a structured plan with a verdict, confidence score, and full transcript.

A separate background worker, the monitor, polls your positions every minute and pings you on Telegram when one drifts out of range or near a boundary - so you know to ask Zuno for a recommendation. See CLI / monitor. The monitor is not a debate participant; it’s a standalone watcher that produces alerts.

$ zuno
◇ recommend what I should do with this position

  ◇ scout       regime: ranging · vol 110bps · gas 14gwei · in range
  ◇ strategist  3 candidates · lean tighter, vol is sleeping
  ◇ critic      reject A: 18h buffer is fine now, vol mean-reverts
  ◇ strategist  A′: ±2.6% · 28h buffer · $3.60/d expected
  ◇ critic      accept A′ - converged
  ────────────────────────────────────────────────────────

  position    WETH / USDC 0.05%
  range       2,190.79 to 2,421.19   out of range
  current     2,438.19

  ────────────────────────────────────────────────────────

  recommended 2,273.36 to 2,614.97   shift
  rejected    2,361.41 to 2,519.99   tighten at 2× vol
  reason      28h base buffer · survives 2× vol · gas/yield 0.71x
  decided by  critic · debate converged

  confidence  0.82  approve_with_caution

Why four agents

LP management is a real tension between fee density (tight ranges, high yield) and survival (wide ranges, low exit risk). A single scoring function picks one side. A debate models the tension explicitly - and makes the reasoning auditable.

agent	does
Scout	reads the position + pool, classifies the regime in plain English
Strategist	proposes 2-5 candidate ranges, revises after critique
Critic	stress-tests every candidate, vetoes weak ones with concrete reasons
Arbiter	only joins on deadlock - picks final candidate from full transcript

Each agent runs as its own AXL peer with its own ed25519 peer id. They talk over Gensyn AXL - the peer-to-peer Agent eXchange Layer. The CLI is just another peer.

A debate normally converges in 1-2 rounds. When it doesn’t, the Arbiter reads the full history and breaks the tie with explicit reasoning the user can read.

What’s deterministic, what isn’t

The agents use the LLM to reason. The LLM never owns arithmetic.

• Tick math, range checks, allocation, and stress simulation are pure functions in TypeScript.
• Volatility comes from a real CoinGecko series; gas comes from eth_gasPrice on the position’s chain; yield is estimated from liquidity × feeTier × in_range_share.
• The Strategist picks (widthMultiplier, centerOffset) proportions - it can’t fabricate raw tick numbers.
• The Critic scores against deterministic floors keyed to the user’s risk profile (conservative | balanced | aggressive).
• The Arbiter writes the final paragraph, but the chosen candidate index is bounded by the actual proposal.

If OPENAI_API_KEY is unset, every agent has a deterministic fallback. The mesh shape and message kinds are identical.

Working surfaces

• Create + rebalance via the four-agent debate over AXL.
• Real on-chain v4 pool discovery (StateView probe of every (token0, token1, feeTier) candidate from the curated registry).
• Live agent_thought transcript rendered in the CLI and persisted in risk.reasons so explain recommendation can replay it.
• In-process orchestrator fallback when the AXL mesh is offline.
• Deterministic fallback when no LLM key is available.
• Background monitor worker with optional Telegram delivery, dedup, and a local ~/.zuno/alerts.json store.
• Email-OTP sign-in, per-user Turnkey sub-org with isolated wallet.
• apply requires explicit approval and signs through the Turnkey boundary only after policy checks pass.