Agent Manage Reference
The agent manage page is your operational control center for a single agent. Access it by clicking Manage on any agent in your dashboard, or navigate directly to /agents/manage?id={agent_id}.
This page is organized into tabs, each covering a different aspect of the agent's configuration and operation.
Settings Tab
The primary configuration tab for your agent.
Agent Details
- Display Name -- The human-readable name shown in dashboards and badges.
- Slug -- URL-safe identifier used in badge URLs (
/v1/badge/{slug}.svg) and public profile links. Unique within your workspace.
- Description -- What this agent does. Displayed on the public agent page.
- Channels -- Which communication channels the agent operates on (web, Slack, email, SMS, etc.).
Widget Embed
Copy-paste embed code for adding a chat widget to your website. The widget connects to the agent's conversation engine and respects all Guard policies.
Configuration options:
- Allowed Origins -- Comma-separated list of domains where the widget can be embedded. Leave empty to allow all origins.
- Theme -- Visual styling of the widget.
LLM Configuration
- Model -- Which LLM provider and model the agent uses (e.g., Claude, GPT-4, Gemini).
- Temperature -- Creativity vs determinism slider.
- System Prompt -- The agent's core instructions and personality (SOUL).
- Max Tokens -- Response length limit.
- Budget -- Monthly spending cap for LLM API calls.
Compliance
- Compliance Framework -- Which regulatory framework applies (HIPAA, SOC 2, GDPR, etc.).
- Data Retention -- How long conversation data is retained.
- PII Handling -- Whether Guard's PII tokenization is active for this agent.
Integrations Tab
Connect third-party services to this specific agent. Each agent can have its own set of integrations independent of other agents in the workspace.
- Available Integrations -- Browse the full catalog of supported services (Google Calendar, Slack, HubSpot, Stripe, etc.).
- Connected Integrations -- View and manage integrations already linked to this agent.
- Per-Integration Settings -- Configure which actions each integration can perform. For example, a Calendar integration might be allowed to read events but not create them.
Integrations use OAuth 2.0, API tokens, or API keys depending on the service. All integration calls flow through Guard for PII tokenization and policy enforcement.
Knowledge Base Tab
Upload documents and crawl URLs to give the agent domain-specific knowledge that it can reference during conversations.
Upload Files
- Drag and drop files or click to browse.
- Supported formats: PDF, TXT, Markdown, DOCX, CSV.
- Files are chunked, embedded, and indexed for retrieval.
Crawl URLs
- Enter a URL to crawl. The system fetches the page content and indexes it.
- Useful for FAQ pages, documentation sites, and product pages.
Document List
- View all indexed documents with their upload date, chunk count, and status.
- Delete documents that are no longer needed.
Knowledge base content is used during the retrieval step of the conversation pipeline. The agent receives relevant chunks as context alongside each user message.
Chat Test Tab
Test your agent in a live conversation interface without deploying it.
- Send messages and see responses in real time.
- The conversation uses the full pipeline: knowledge retrieval, Guard PII tokenization, LLM routing, integration execution, and grounding checks.
- Useful for validating prompt changes, knowledge base updates, and integration configurations before going live.
- Each test session creates a real conversation with a conversation ID that appears in the conversation logs.
Conversation Logs Tab
View the history of all conversations this agent has participated in.
- Conversation List -- All conversations sorted by recency, with conversation IDs, channel, turn count, and timestamps.
- Conversation Detail -- Click into a conversation to see the full message exchange, including user messages, agent responses, tool calls, and integration actions.
- Filtering -- Filter by channel, date range, or conversation ID.
Conversation logs are retained according to your workspace's data retention policy.
Vault Tab
The immutable audit ledger for this agent. Available on the Max plan.
- Event Ledger -- Every significant event (score changes, decisions, tool calls, policy violations) is recorded with a hash-chain that proves integrity.
- Chain Verification -- Verify that the audit chain has not been tampered with. Each entry includes a hash of the previous entry.
- Export -- Download the audit ledger as JSON for external compliance systems.
See the Vault documentation for details on chain verification and evidence integrity.
Passport Tab
Agent identity verification and portable credentials. Available on Pro+ plans.
- Verification Status -- Whether the agent's identity has been verified through key attestation, domain verification, or manifest signing.
- Manifests -- View and manage the agent's signed manifest, which declares its capabilities, permissions, and identity claims.
- Delegations -- If this agent delegates tasks to other agents, manage delegation chains and trust relationships here.
- Portable Credentials -- Issue and manage ES256 JWT credentials that other platforms can verify via your JWKS endpoint.
See the Passport documentation and Portable Credentials for details.
Security Tab
Guard security scanning and operational controls. Available on the Max plan.
Guard Scan
Run on-demand security scans against the agent's configuration:
- SOUL/Prompt Injection Scan -- Checks the system prompt for injection vulnerabilities, bypass attempts, and content policy violations.
- MCP Tool Scan -- Analyzes connected MCP tools for security risks.
- Findings List -- View all past scan findings with severity levels and remediation guidance.
Red-Team Tests
Automated adversarial testing that attempts to:
- Extract the system prompt
- Bypass content policies
- Trigger unintended tool calls
- Exploit integration permissions
Results show which attacks succeeded and specific recommendations for hardening.
Kill Switch
Emergency stop for the agent. When activated:
- All trust decisions for this agent return
deny.
- The agent cannot process new conversations.
- Existing conversations are terminated.
- A
agent.killed webhook event is dispatched.
The kill switch can be deactivated to restore normal operation (agent.unkilled event).
Benchmarks
Performance and quality benchmarks for the agent:
- Response latency percentiles
- Token usage statistics
- Grounding accuracy (how well responses align with knowledge base content)
- Cost per conversation
See the Guard documentation for details on security scanning and policy enforcement.