The Agent Loop

The agent loop is a structured 6-step cycle for AI-powered capital management on Gearbox. Each stage has clear inputs, outputs, and decision points.

Step 1: Discover

Scan all available opportunities across chains. This is a broad survey — every pool, every strategy, every chain. The data is lightweight: headline APY, TVL, access requirements.

const opportunities = await sdk.opportunities.search({
  chainIds: ["Mainnet", "Monad"],
  types: ["pool", "strategy"],
  assets: [Asset.STABLE],
});

The agent applies its own filters (APY floor, TVL minimum, permissionless-only, etc.) and produces a shortlist of candidates for deeper analysis.

Step 2: Analyze

Due diligence on the shortlisted candidates. The agent inspects each one in detail using specialized research sub-agents:

// Deep dive into a candidate
const detail = await sdk.pools.getDetail({ chainId, address: poolAddress });
const curator = await sdk.curators.getProfile(detail.curatorId);
const apyHistory = await sdk.history.getMetric({ target, metric: "apy", periodDays: 30 });
const events = await sdk.events.getFeed({ chainId, target, sinceDays: 30 });

The output is AnalyzedOpportunity[] — ranked by score, each with profitability forecasts, risk breakdowns, and reasoning.

Step 3: Propose

From the analyzed candidates, find the optimal action — or decide that no rebalance is needed.

The agent considers:

• Is the current position already optimal?
• Would the rebalance cost (gas, slippage) exceed the expected gain?
• What's the best route for the chosen strategy?

If a rebalance makes sense:

// Query router for optimal path
const route = await sdk.router.findOpenStrategyPath({
  chainId,
  creditManagerAddress,
  collateralToken,
  collateralAmount,
  debtAmount,
  targetToken,
  slippageBps: 50,
});

// Build the unsigned transaction
const tx = await sdk.accounts.openCreditAccount({
  chainId,
  creditManagerAddress,
  collateralToken,
  collateralAmount,
  debtAmount,
  pathCalls: route.calls,
  slippageBps: 50,
});

If no action is needed, the agent skips to Monitor.

Step 4: Preview

Verify that the transaction is feasible on-chain right now. LLMs are not instant — by the time the agent finishes reasoning, on-chain conditions may have changed (liquidity exhausted, price moved, quota filled).

const preview = await sdk.previewTransaction(tx);

The preview simulates the exact bytes against current chain state and returns:

• success — would this execute?
• healthFactor — projected HF after entry
• actions[] — human-readable step descriptions
• balanceChanges[] — net token movements
• warnings[] — concerns (stale oracle, high utilization, etc.)

If preview fails → back to Propose

The agent does not re-analyze. The due diligence is still valid — the candidates are sound. Only the execution parameters need adjustment:

• Liquidity exhausted → reduce position size or try different route
• Price moved → adjust slippage tolerance
• Quota filled → switch to the next-ranked candidate from Analyze

if (!preview.success || (preview.healthFactor ?? 0) < 1.4) {
  // Adjust parameters and re-propose
  // Do NOT re-run Analyze — the research is still valid
}

Step 5: Execute

Sign and submit the transaction. The SDK never signs — execution is the caller's responsibility.

Two paths:

• Human-in-the-Loop — agent encodes preview as URL for verify.gearbox.finance, human reviews and signs
• Bot Execution — agent signs via bot API with bounded on-chain permissions

const txHash = await walletClient.sendTransaction({
  to: tx.to,
  data: tx.calldata,
  value: tx.value,
});

Step 6: Monitor

Watch the live position and react to events. The monitor waits for signals — either scheduled (cron) or event-driven (alerts, external data).

const position = await sdk.accounts.getStatus({ chainId, creditAccount });
const alerts = await sdk.monitor.getAlerts({ chainId, creditAccount });

When events trigger a review → back to Analyze

The monitor doesn't jump to Propose — it goes back to Analyze to re-evaluate the situation with fresh data. This is critical: if a collateral token is dropping, the agent needs to understand the severity before deciding what to do.

The urgency of the trigger determines the pace. A cron check runs the full cycle. A critical event skips straight to action — but still passes through Analyze to confirm the threat is real before executing.

Learn More

• MCP Server — MCP tools organized by loop stage
• Execution — How agents interact with the protocol safely