Skip to main content

AI Agents Thesis

AI Agents Investment Appraisal

How do you evaluate tokens for autonomous agents that trade, yield-farm, and make decisions on-chain?


The Thesis

Autonomous AI agents will capture a significant share of blockchain transaction value through automated execution, yield optimization, and intent-based routing. The core bet: on-chain agents outperform manual DeFi interaction on speed, consistency, and risk management.

Why It Matters

Speed and consistency. Agents execute 24/7. No sleep, no emotion, no missed opportunities. Arbitrage windows that last seconds become accessible.

Composability. DeFi protocols are permissionless Lego blocks. Agents can chain actions across protocols faster than any human — rebalance, harvest, compound, bridge, hedge in a single transaction.

Intent alignment. Users define goals ("maximize yield at below 5% drawdown risk"). Agents find the best path. The shift from manual execution to intent-based delegation.

Cost compression. Agent-executed strategies reduce the cost of active portfolio management from basis points to gas fees.

Where It Applies

CategoryWhat They DoExample Tokens
DeFi Yield AgentsAutomated yield farming, LP management, harvest-compound loopsFET, VIRTUAL
Arbitrage AgentsCross-protocol and cross-chain arbitrage executionAI16Z
VC AutomationDeal flow analysis, due diligence, portfolio monitoringFAI, TMAI
Wallet AgentsPortfolio rebalancing, tax-loss harvesting, risk alertsElizaOS ecosystem
Trading AgentsMarket making, momentum strategies, sentiment-based executionVarious DeFAI tokens

Market Sizing

  • DeFi TVL fluctuates between $50-200B — agents optimizing even 10% of this creates massive value
  • On-chain transaction value growing as institutional adoption increases
  • Agent-executed transactions as percentage of total on-chain volume is the key metric to watch

Evaluation Checklist

1. Agent Capability (Weight: 25%)

  • What specific tasks does the agent perform? (Not just "AI-powered")
  • Task completion rate — does it actually work? Verifiable on-chain
  • Autonomous action frequency — how often does it act without human intervention?
  • Inference cost per action — is the AI cost structure sustainable at scale?
  • Failure recovery — what happens when the agent makes a bad decision?

2. Token Economics (Weight: 25%)

  • Token utility — does holding/staking the token actually improve agent performance?
  • Value capture — how does protocol revenue flow to token holders?
  • Emission schedule — inflationary or deflationary? Vesting timeline for team/investors
  • Usage growth vs dilution — does adoption outpace token inflation?
  • Incentive alignment — are agent operators rewarded for performance or just participation?

3. Technical Architecture (Weight: 20%)

  • On-chain vs off-chain execution — what runs on-chain? What's just an API call?
  • Smart contract audits completed before mainnet
  • Open source model weights or proprietary? Verifiable inference?
  • Composability — does it integrate with existing DeFi protocols or require its own ecosystem?
  • Upgrade mechanism — can the AI model be improved without breaking positions?

4. Team and Governance (Weight: 15%)

  • Team credentials in both AI/ML and blockchain
  • Prior delivery — shipped products, not just papers
  • Governance mechanism — who decides model updates?
  • Transparency — are agent decisions auditable? Can users understand why an action was taken?

5. Market Position (Weight: 15%)

  • Competitive moat — what prevents a better model from forking the protocol?
  • Network effects — does the agent improve as more users join?
  • Data advantage — does the protocol accumulate proprietary data that improves performance?
  • Integration partnerships — which DeFi protocols and wallets support it?

Scorecard

DimensionScore (1-5)Weight
Agent Capability25%
Token Economics25%
Technical Architecture20%
Team and Governance15%
Market Position15%
AGGREGATE100%

Conviction Mapping

AggregateConvictionTypical Allocation
4.0+5/5 — Proven agent with live revenueUp to 7%
3.5-4.04/5 — Strong traction, clear moatUp to 5%
3.0-3.53/5 — Working product, early adoption2-3%
2.0-3.02/5 — Promising tech, unproven market fit1-2%
Below 2.01/5 — Exploratory0.5-1%

Risk Factors

AI agent tokens carry unique risks beyond standard crypto exposure:

  • Model risk — AI makes a bad trade and the loss cascades. Smart contract bugs combined with AI autonomy = amplified damage
  • Regulatory risk — Autonomous financial agents may trigger securities or investment advisor regulations
  • Narrative collapse — "AI" prefix inflates valuations. When the hype cycle turns, fundamentals-free projects crash hardest
  • Inference cost compression — Model efficiency gains make proprietary AI less defensible over time
  • Oracle dependence — Agents relying on external data feeds inherit all oracle attack vectors
  • Black box trust — Users delegate capital to systems they can't fully audit. One bad update away from loss

Validation Metrics

Track these monthly:

  • AI agent transaction count growth (target: >12% monthly)
  • Protocol-owned agent annualized revenue (target: >$50M for top protocols)
  • Percentage of top 100 DeFi protocols integrating AI execution layers (target: >40%)
  • Agent-executed volume as share of total DEX volume
  • Average agent ROI vs passive hold benchmark

Allocation Balance

CategoryExample AssetsRationale
Agent ProtocolsFET, VIRTUAL, TMAIInfrastructure for autonomous ops
AI-Powered VCAI16Z, FAIAutomated deal flow analysis
Agent FuelETH, SOL, TIATransaction/gas fee beneficiaries
Stable YieldUSDe, sDAIRisk mitigation during volatility

Guardrail: Automatic rebalancing when AI agent category exceeds 40% of portfolio. Protocol-level stop losses at -25% from entry. Quarterly review of agent performance vs ETH benchmark.

Context

Questions

If AI agents can execute faster and more consistently than humans, what remains of the human investor's edge?

  • When an agent's inference costs drop 10x, does the protocol's token retain value — or does the moat disappear with the cost?
  • How do you audit an autonomous agent's decision-making when the model is a black box?
  • At what point does "AI-powered" stop being a valid premium and start being a warning sign of narrative over substance?