Agent Guardrails Docs

Dashboard Guide

The Agent Guardrails dashboard provides real-time visibility and control over your autonomous agents.

Authentication

The dashboard uses SIWS (Sign In With Solana) for authentication. Connect your Phantom, Solflare, or Backpack wallet, then sign a nonce message to prove ownership. A JWT is issued as an httpOnly cookie and all subsequent data is filtered to policies owned by your wallet only.

Fleet Home

The fleet overview page gives you a snapshot of your entire agent fleet at a glance:

  • Active agent count— agents currently running with active policies
  • Paused agent count— agents that have been frozen by the kill switch or monitoring pipeline
  • Incidents in last 24h— pause events with trend indicators
  • Total SOL spent in 24h— aggregate spend across all agents

Quick access links take you directly to the full agent list or recent incidents.

Creating a Policy

The 4-step policy creation wizard walks you through configuring a new agent policy:

  1. Select Programs— choose from Jupiter, Token Program, System Program, or enter custom program IDs. Up to 10 programs can be added to the allow-list.
  2. Set Limits— configure the per-transaction SOL cap and daily SOL budget. Values are entered in SOL and automatically converted to lamports.
  3. Session Expiry— set the number of days from now until the session key expires. This is converted to a Unix timestamp on-chain.
  4. Squads Escalation (optional)— configure a Squads multisig address and threshold amount. Transactions exceeding this threshold are escalated to the multisig for approval.

Agent Detail Page

The agent detail page provides comprehensive monitoring for a single agent:

  • Spend Gauge— a radial chart showing daily SOL spent versus the configured budget
  • Activity Feed— live SSE-updated transaction list showing all recent activity
  • Policy Summary Card— displays current limits, allowed programs, and status
  • Controls— kill switch, edit policy, rotate key, fund agent, and close policy actions

Kill Switch

Click the kill switch button to immediately pause an agent. A confirmation modal appears, and upon confirming, the dashboard signs a pause_agent instruction on-chain. The agent is frozen immediately and cannot execute any further transactions until resumed by the policy owner.

Pausing an agent is immediate and on-chain. The agent cannot execute any further transactions until resumed.

Activity Feed

The activity feed is a real-time transaction stream powered by Server-Sent Events (SSE). Each row displays:

  • Transaction signature
  • Target program
  • Amount
  • Verdict badge — ALLOW (green), FLAG (amber), PAUSE (red)
  • Confidence score
  • Timestamp

Click any row to expand it and view the full verdict reasoning provided by the Guardian Agent.

Incidents

The incidents page lists all agent pauses across your fleet. Each incident shows:

  • Pause time
  • Who paused— the policy owner or an authorized monitor
  • Reason for the pause
  • Triggering transaction

Click an incident for the full detail view, which includes:

  • Incident timeline— a vertical timeline of all events leading up to and following the pause
  • Guardian Agent postmortem— a detailed markdown report generated by the AI judge explaining the anomaly
  • Judge verdict chain— the sequence of verdicts that led to the pause decision

Escalations

When a transaction exceeds the Squads escalation threshold, the following flow occurs:

  1. The guardrails program rejects the transaction with an EscalatedToMultisig error
  2. The dashboard shows an escalation notification
  3. A Squads proposal is created for multisig review
  4. Multisig members approve the proposal
  5. The transaction is executed via the multisig_execute instruction

Playground

The attack simulator is a demo and testing tool for understanding how the monitoring pipeline responds to different transaction patterns. In the playground you can:

  • Configure agents with custom policies
  • Craft transactions with specific parameters
  • Inspect prefilter signals as they fire
  • View full verdict reasoning from the Guardian Agent

This is useful for validating your policy configuration before deploying to production, and for understanding why specific transactions are flagged or paused.