Personality & Guardrails

Personality and guardrails form the core behavioral foundation of every agent. While personality (or persona) defines how an agent communicates and approaches problems, guardrails define the strict boundaries of what it is allowed to do.

Together, these settings ensure that an agent remains helpful, on-brand, and aligned with company security policies rather than becoming an unmonitored operational liability.

Defining agent personality

An agent’s personality shapes user interactions and sets expectations for how it handles tasks. Personality is configured using three primary components:

• Tone — the communication style the agent uses in chat and generated reports (for example, direct and concise, formal and analytical, or warm and supportive).
• Working style — how the agent approaches problem-solving and task execution (for example, proactive: proposes a multi-step plan before executing, or methodical: asks clarifying questions when parameters are missing).
• Values and rules — core principles the agent adheres to during execution (for example, always cites source documents for factual claims or highlights uncertain assumptions).

When an agent is published to the internal marketplace, these personality details appear on its listing card so team members know exactly how it behaves before adopting it into their workspace. See Agent marketplace.

Action-level guardrails and permissions

Guardrails assign explicit permission levels to specific system actions, tools, and data sources:

• Allowed — the agent executes the action freely without requiring user intervention (ideal for low-risk, read-only tasks like searching internal docs or summarizing text).
• Needs approval — the agent pauses execution and submits an approval request to a human team member before proceeding (used for high-impact actions like updating a CRM, sending external emails, or running financial transactions). See Approval workflows.
• Blocked — the agent refuses to perform the action under any circumstances (used to prevent dangerous operations like modifying production databases or exposing customer personally identifiable information).

Which actions a given person can set, and who can change them, is defined by role. See Capability matrix.

Organization-wide enterprise guardrails

In an enterprise deployment, administrative policies sit above individual agent configurations:

• System-enforced safety floor — IT administrators establish global guardrails that apply automatically across every agent and workspace in the organization.
• Non-bypassable protection — individual agent builders cannot override or disable organization-level guardrails. Even if an employee creates a custom agent and forgets to block sensitive operations, global policies prevent unauthorized actions automatically.
• Approved model enforcements — global guardrails dictate which foundation language models and endpoints can be used, preventing unapproved third-party model routing.

For how administrators configure these, see Guardrails and allowed models.

How guardrails work during scheduled runs

When an agent executes autonomously on a recurring schedule (a ritual), there is no human active in the chat window to grant real-time sign-offs.

• Handling approval steps — any action marked Needs approval is automatically skipped during an autonomous scheduled run, and a note is recorded in the execution log.
• Configuring scheduled agents — for a ritual to complete its tasks without interruption, all required actions for that scheduled run must be explicitly set to Allowed within established security policies.

Check permissions before scheduling

An action left at Needs approval will be skipped silently during a ritual, not queued for later sign-off. Review every action a scheduled run depends on.

Example: customer account assistant

Consider an analyst who builds an agent to assist with customer account lookups:

1. Personality setup — the agent’s tone is set to concise and factual, and its values require it to always quote verified account records.
2. Agent-level guardrails — the analyst sets reading account summaries to Allowed, but sets exporting full customer data files to Blocked.
3. Organization-level guardrails — sitting above the analyst’s settings, company-wide enterprise policies forbid any agent from transmitting data to unapproved external endpoints.

The agent operates safely by default, fully compliant with company standards without requiring the individual builder to anticipate every potential security risk.

Guardrail permission summary

Permission level	Execution behavior	Best used for	Behavior during rituals
Allowed	Executes automatically	Read-only tasks, data lookups, draft reports	Runs smoothly without stopping
Needs approval	Pauses for human sign-off	Sending emails, updating CRM records, making payouts	Skipped automatically with a log entry
Blocked	Refuses action completely	Deleting records, sharing PII, using unapproved tools	Refused completely

Next steps

• Set up autonomous execution schedules → Rituals
• Learn how administrators configure company-wide rules → Guardrails and allowed models
• Explore full agent architecture → What is a Rival Agent