The problem MandateAgent solves

Trading on-chain options requires users to complete several complex tasks at the same time: understand market conditions, compare strategies, find live liquidity, calculate potential loss, verify contract details, manage token approvals, and submit a transaction before the quote expires.
MandateAgent makes this workflow easier and safer by:

• Turning a natural-language trading objective into clear, reviewable rules, including strategy type, expiry range, spending limit, maximum fills, and prohibited actions.
• Combining live ETH prices, volatility indicators, market news, public on-chain activity, and Thetanuts liquidity in one command center.
• Using a committee of specialist AI agents to research the market, recommend a strategy, and independently verify the recommendation.
• Showing evidence sources, confidence scores, payoff outcomes, policy checks, and agent progress so users understand exactly what they are approving.
• Enforcing deterministic safety controls that AI cannot override, including a maximum 25 USDC cumulative budget, quote freshness, contract allowlists, strike ordering, wallet balances, and an emergency stop.
• Preparing live Thetanuts OptionBook transactions on Base mainnet through MetaMask, with a bounded Coinbase CDP wallet path for autonomous execution.
• Recording every prepared, submitted, confirmed, failed, or cancelled execution in an auditable trade ledger.
  If the evidence is unreliable, the quote is stale, liquidity is unavailable, or any rule fails, MandateAgent returns HOLD and refuses to prepare the transaction.

Challenges we ran into

Keeping live options quotes valid
Thetanuts orders can change or disappear between market analysis and wallet confirmation. We addressed this by refreshing the selected order immediately before execution. The transaction is rejected if its premium, strikes, expiry, quantity, venue, or contract differs from the approved proposal.
Making AI safe enough for wallet actions
AI recommendations are probabilistic, while financial limits must be deterministic. We separated the AI committee from the policy engine. Agents can recommend a strategy, but they cannot increase the budget, approve an unverified contract, bypass the verifier, ignore an expired mandate, or disable the emergency stop.
Preventing duplicate spending
Manual MetaMask execution and autonomous execution could attempt to use the same mandate budget concurrently. We introduced atomic budget reservations and idempotency keys. Pending, submitted, and confirmed transactions all count toward the cumulative spending ceiling.
Explaining what the agents are doing
An early interface only showed that the committee was “waiting,” which made it difficult to understand its progress. We added named agent stages, current tasks, completion percentages, evidence summaries, sources, timestamps, reliability labels, and verifier conclusions.
Handling incomplete data and unavailable liquidity
News services, model providers, market APIs, and OptionBook liquidity can fail independently. Each source is time-bounded and scored for freshness and reliability. Missing evidence, model disagreement, or a lack of qualifying liquidity produces HOLD instead of a fabricated trade opportunity.
Building reliable transaction tracking
A transaction request can time out even when it was submitted successfully, so blindly retrying could create a duplicate trade. MandateAgent stores transaction hashes and reconciles receipts before deciding whether another submission is safe. Every lifecycle change is written to an append-only execution ledger.
Integrating changing on-chain SDKs
Thetanuts orders use signed structures, contract-specific units, and live allowlists. We used the official client for discovery, previews, and calldata encoding, normalized token decimals explicitly, and kept OptionFactory settlement fail-closed until its production parameters can be independently verified.