Key concepts in Project Perception

Important

Some information in this article relates to a prereleased product which may be substantially modified before it's commercially released. Microsoft makes no warranties, expressed or implied, with respect to the information provided here.

Project Perception uses the following concepts to organize agentic work: sessions, agents, and playbooks. Understanding these concepts helps you configure, delegate, and supervise agentic operations effectively.

Sessions

A session is the record of an agent interaction. Sessions capture agent activity, decisions, inputs, outputs, and any approvals or human interventions. Sessions can involve a single agent or multiple agents working in coordination. Each session has a status that reflects the aggregate state of all tasks within it.

For more information about sessions, see View and manage sessions.

Agents

An agent is an autonomous or semi-autonomous software entity that can reason, use tools, and perform tasks. Project Perception agents are Microsoft-published agents designed for deep, autonomous security operations work. They form the core of Microsoft's multi-agent SOC system, covering the full security operations lifecycle from threat intelligence through detection authoring, exposure analysis, posture prioritization, alert triage, and incident investigation.

Perception agents are designed to collaborate. The output from one agent becomes input for another, enabling end-to-end workflows. For example, threat intelligence extracted by the Threat Intelligence Agent feeds directly into detections created by the Detection Authoring Agent.

Perception agents require an agent identity and role assignment before they can be used. For setup instructions, see Work with agents.

Perception agents include:

Agent	Agent team	Description
Recon Agent	Red team	Performs attacker scouting activities to map out your environment. Use this agent to surface attack paths, choke points, valuable assets, configuration issues, and excess permissions.
Triage Agent	Blue team	Classifies and resolves various types of alerts as true or false positives. This agent provides a detailed explanation behind its conclusions, learning from context and user feedback to improve accuracy.
Threat Intelligence Agent	Blue team	Extracts attack patterns, indicators, and other key TI objects and analyzes them for insights on how to defend against or respond to a threat.
Attack Investigation Agent	Blue team	Reconstructs the full attack story across alerts, incidents, and other security signals. This agent provides analyst-ready insights, including verdicts, timelines, attack techniques, indicators of compromise, and affected entities.
Detection Authoring Agent	Blue team	Identifies gaps in threat coverage and creates detection rules to enhance defenses as threats evolve.
Posture Prioritization Agent	Green team	Ranks posture issues broadly or in the context of specific threats by real-world risk, exposure, exploitability, and asset criticality. This agent provides clear, focused guidance on what to fix first and why.

Playbook

A playbook is a reusable, agent-driven set of actions for a specific security objective. When run, agents execute the playbook's steps on behalf of the user, investigating signals, correlating data, and producing output without requiring manual intervention at each step.

Key characteristics of playbooks:

• Preset by Microsoft. The playbook list is read-only. You can't create custom playbooks in this release.
• Can use multiple agents. A single playbook can involve multiple agents, and a single agent can be used by multiple playbooks.
• Might require an input to run. Depending on the playbook, a session might not start until you provide the specific input it needs. For example, an Investigate incident playbook requires an incident ID before it can run.
• Require all agents to be enabled. A playbook is only active when all of its required agents are set up and enabled.

Perception includes the following playbooks. Some playbooks run a single agent for a focused task. Others coordinate multiple agents in sequence, where each agent's output feeds the next.

Playbook	Description	Participating agents
Extract threat intelligence	Analyzes a threat intelligence article and produces structured intelligence objects, including TTPs, IOCs, CVEs, and KQL hunting queries.	Threat Intelligence Agent
Protect against a threat	Multi-agent workflow that takes a threat intelligence source and produces attack path analysis, posture findings, prioritized remediation, and new detection rules.	Threat Intelligence Agent, Recon Agent, Posture Prioritization Agent, Detection Authoring Agent
Investigate incident	Performs an autonomous tier-2 investigation of a Defender incident and produces a complete attack story with verdict, timeline, and remediation actions.	Attack Investigation Agent
Identify attack paths	Maps how an attacker could move through your Azure environment and reach valuable assets, with choke-point analysis and remediation recommendations.	Recon Agent
Assess identity risks	Evaluates identity exposure and lateral movement risks in your Azure environment, with identity-focused attack path analysis and remediation guidance.	Recon Agent
Plan for posture remediation	Build a prioritized remediation plan across clouds, devices, and AI. This playbook weighs exposure, exploitability, and asset context to target the highest-risk posture gaps.	Posture Prioritization Agent

How the concepts relate

The following table shows how the concepts work together:

Concept	Creates or uses	Example
Sessions	Contains agent work; can be created by playbooks	A session investigating a phishing incident involves three agents working in sequence.
Agents	Participates in sessions	A threat intelligence agent joins a session to research indicators of compromise.
Playbooks	Creates sessions; assigns agents	A playbook for high-severity incidents creates a session with triage, investigation, and response agents.

Next steps

• Get started with Project Perception
• Work with agents