Execution

How agents interact with the Gearbox protocol on-chain — and why it's safe.

The Security Model

A core principle of Gearbox Agentic: the SDK builds transactions, but never signs them. Transaction construction (what to do) is separated from transaction execution (actually doing it). This creates a clear security boundary.

Every transaction goes through preview before execution:

1. Build — sdk.positions.prepareOpen() → produces a RawTx { to, calldata, value }
2. Preview — sdk.previewTransaction(rawTx) → simulates the exact bytes, returns success, health factor, actions, balance changes, warnings
3. Validate — agent (or human) checks: success = true, HF > threshold, no critical warnings
4. Execute — sign and send via wallet

The same bytes that were previewed are the bytes that go on-chain. No deviation.

Two Execution Modes

Mode	Trust Level	Best For
Human-in-the-Loop	Agent proposes, human approves	High-value positions, institutional compliance, initial trust building
Bot Execution	Agent executes autonomously within bounds	Rebalancing, liquidation monitoring, automated management

Both modes use the same preview mechanism. The difference is who signs.

Learn More

• Human-in-the-Loop — verify.gearbox.finance approval flow
• Bot Execution — bounded on-chain permissions
• The Agent Loop — how Preview and Execute fit in the 6-step cycle