Guardrails & Policies

Define behavioral rules for your agents, enforce compliance, and intervene when things go wrong.

Overview

The Guardrails page is your central control panel for agent safety. It lets you define policies that agents must follow, track whether agents have acknowledged those policies, inspect agent memory, and set up automated rules that trigger interventions (pause, kill, alert) when agents misbehave.

Think of it as the compliance layer between your organization's rules and your AI agents' behavior. Policies are written in plain language (Markdown), and agents internalize them during an onboarding flow before they start operating.

Dashboard Tabs

Tab	Purpose
Overview	Quick stats — policy count, onboarding progress, recent interventions. Also has agent pause/resume/kill controls.
Policies	Create, edit, and manage behavioral policies grouped by category.
Onboarding	Track which agents have acknowledged your policies. Force-complete or re-initiate onboarding.
Agent Memory	View and manage what agents remember — org knowledge, corrections, preferences, and more.
Rules & Interventions	Configure automated guardrail rules, anomaly detection thresholds, and review the intervention log.

Start here: The Overview tab gives you a quick health check. If everything looks green, you're good. Drill into specific tabs for detailed management.

Creating & Managing Policies

Policies are the rules your agents must follow. Each policy has a name, category, enforcement level, and content written in Markdown.

Policy Categories

Category	Use For
Code of Conduct	General behavioral expectations — tone, professionalism, boundaries
Communication	Rules about how agents communicate — response style, escalation triggers, language
Data Handling	PII protection, data retention, what agents can and cannot store
Brand Voice	Tone, terminology, messaging guidelines
Security	Credential handling, API usage limits, prohibited actions
Escalation	When and how agents should escalate to humans
Custom	Anything else specific to your organization

Creating a Policy

Go to the Policies tab and click Create Policy
Enter a Name and choose a Category
Set the Enforcement level (see below)
Write the policy Content in Markdown — be specific and actionable
Set Applies To — use * for all agents, or list specific agent IDs
Set Priority (0 = highest priority, shown first)
Click Create Policy

Default Templates: Click Apply Default Templates to load a starter set of policies covering common scenarios like data handling, communication standards, and escalation procedures.

Policy Content Tips

Write policies as clear, actionable instructions. Agents internalize these during onboarding and store them in memory. Good policies are:

Specific: "Never share API keys in email" beats "Handle credentials carefully"
Actionable: Tell the agent what to do, not just what not to do
Testable: Could you check if the agent followed this? If not, make it more concrete

Enforcement Levels

Level	Behavior	Use When
Mandatory	Agent must acknowledge before operating. Violations can trigger interventions.	Security policies, legal compliance, data handling rules
Recommended	Agent should follow but won't be blocked. Flagged in audits.	Best practices, style guides, preferred workflows
Informational	Provided for context only. No enforcement.	Background info, org culture notes, nice-to-knows

Versioning: When you update a mandatory policy, agents that already acknowledged the old version may need re-onboarding. Use the Check Policy Changes button in the Onboarding tab to detect stale acknowledgments.

Agent Onboarding

Onboarding is the process by which agents acknowledge and internalize your organization's policies. Until an agent completes onboarding, it operates with limited capabilities.

Onboarding Statuses

Status	Meaning
Completed	Agent has acknowledged all required policies
In Progress	Agent has acknowledged some policies but not all
Needs Renewal	Policies changed since last acknowledgment — re-onboarding needed
Not Started	Agent hasn't begun onboarding

Managing Onboarding

Initiate Onboarding: Select an agent and click to start the process
Force Complete: Skip the acknowledgment flow and mark an agent as onboarded (use sparingly)
Check Policy Changes: Scan for agents whose onboarding is stale due to policy updates

Agent Memory

The Agent Memory tab lets you inspect, create, and manage what agents remember. Memories are categorized and have confidence scores that decay over time.

Memory Categories

Category	Examples
Org Knowledge	Company facts, team structure, product details
Interaction Pattern	How a specific user prefers to communicate
Preference	User likes morning summaries, prefers bullet points
Correction	"The CEO's name is spelled Kathryn, not Katherine"
Skill	Learned capabilities from training or experience
Context	Background information for ongoing projects
Reflection	Agent's self-assessment and lessons learned

Memory Management

Add Memory: Manually inject knowledge into an agent's memory (e.g., "Our Q3 revenue target is $2M")
Prune: Remove expired memory entries
Decay: Reduce confidence scores on old memories, making them less influential
Delete: Remove specific memory entries

Confidence scores range from 0–100%. Memories above 70% are considered reliable. Below 30%, they're treated as uncertain. Use the Decay button to gradually reduce confidence on stale information.

Rules & Automated Interventions

Rules let you define automated responses to agent behavior. When a rule's conditions are met, the system automatically takes action — alerting admins, pausing agents, or killing sessions.

Rule Categories

Category	What It Detects
Anomaly Detection	Unusual patterns — excessive API calls, error spikes, abnormal token usage
Policy Compliance	Violations of defined policies
Communication	Inappropriate messaging, tone violations
Security	Credential exposure, unauthorized access attempts

Rule Actions

Action	Effect
Alert	Send a notification to admins
Pause Agent	Temporarily suspend the agent — can be resumed from the dashboard
Kill Agent	Immediately terminate the agent session
Notify Admin	Send a notification without affecting the agent
Log Only	Record the event for auditing but take no action

Severity Levels

Each rule has a severity (Low, Medium, High, Critical) that determines how urgently the intervention is surfaced in the dashboard and notifications.

Manual Interventions

From the Overview tab, you can manually control any agent:

Pause: Temporarily suspends the agent. It stops processing but retains state.
Resume: Resumes a paused agent.
Kill: Immediately terminates the agent session. Use for emergencies.

All interventions (manual and automated) are logged in the intervention history with timestamps, the triggering user or rule, and the reason.

Kill is irreversible for the current session. The agent's session is terminated immediately. You'll need to restart the agent from the Agents page.

Best Practices

Start with default templates — they cover common scenarios and give you a solid foundation
Use mandatory enforcement sparingly — reserve it for security and compliance. Over-restricting agents reduces their effectiveness
Review the intervention log weekly — patterns in interventions reveal whether your policies need tuning
Keep policies concise — agents internalize every word. Long, vague policies dilute important rules
Set up anomaly rules early — catch runaway agents before they cause damage
Use cooldown periods on rules to prevent alert fatigue from repeated triggers
Decay memory periodically — old information becomes less reliable. Let confidence scores reflect that

Troubleshooting

Issue	Solution
Agent not following policies	Check the Onboarding tab — the agent may not have acknowledged the policy. Force-complete or re-initiate onboarding.
Too many intervention alerts	Increase cooldown periods on rules, or change severity from "alert" to "log only" for low-priority rules.
"Needs Renewal" showing for all agents	You recently updated a mandatory policy. Click Check Policy Changes to re-onboard affected agents.
Memory entries not appearing	Select the correct agent from the dropdown. Memories are per-agent. Also check category and importance filters.
Paused agent won't resume	Check if the agent was killed (not paused). Killed agents need to be restarted from the Agents page.

Audit Log — Review all administrative actions and policy changes
Multi-Tenant Guide — Organization-scoped policies and agent assignment
Settings — Configure security, tool permissions, and network rules

AgenticMail Enterprise — Guardrails & Policies Documentation

AgenticMail Enterprise Documentation Report an issue

Guardrails & Policies

On This Page

Overview

Dashboard Tabs

Creating & Managing Policies

Policy Categories

Creating a Policy

Policy Content Tips

Enforcement Levels

Agent Onboarding

Onboarding Statuses

Managing Onboarding

Agent Memory

Memory Categories

Memory Management

Rules & Automated Interventions

Rule Categories

Rule Actions

Severity Levels

Manual Interventions

Best Practices

Troubleshooting

Related Pages