Agent Loop
How each Sigil Wallet agent reasons, validates, and executes on a continuous loop.
Every Sigil Wallet agent runs an independent, continuous loop powered by LangGraph. Each cycle follows a strict graph of nodes with conditional edges.
The Loop Cycle
Node Details
1. Gather Context
The agent fetches its current state from Solana Devnet:
- SOL balance
- SPL token holdings and associated accounts
- Recent transaction history
2. Agent Reasoning
The gathered context is sent to the configured LLM. The model evaluates whether any action is needed.
If the model determines no action is required, the loop sleeps until the next cycle (default: 60 seconds).
If an action is needed, the model generates a tool call — a structured intent like transfer_sol({ to: "...", amount: 0.5 }).
3. Guardrail Validation
Every proposed tool call passes through the Guardrails layer before execution. This layer is hard-coded and operates independently of the LLM.
Checks include:
- Per-trade value limits
- Daily volume caps
- Slippage tolerance for swaps
- Recipient allowlists
- Cool-down period enforcement
If validation fails, the action is logged with a reason and skipped.
4. Execute Tool
Approved actions are routed to the Wallet Layer for execution:
- Key is retrieved from the OS Keychain via
keytar - Transaction is built using
@solana/web3.js - Transaction is signed with the agent's keypair
- Signed transaction is submitted to Solana Devnet
- Result (signature or error) is logged to the database
5. Loop Back
After execution (or skip), the agent returns to the gather_context node and the cycle repeats.
Configuration
| Setting | Default | CLI Command |
|---|---|---|
| Loop interval | 60 seconds | sigil config set --loop-interval <seconds> |
| Max retries | 3 | Configurable per agent |
Key Principle
The agent loop is stateless between cycles. Each iteration starts fresh by gathering the latest on-chain data. This prevents stale state issues and ensures the agent always reasons about current reality.