Agents
BOSS agents are named AI team members with identity, memory, and accountability. They are not anonymous chatbots -- they build trust over time and operate within governance boundaries.
Named agents with identity
Every BOSS agent has a name, a defined role, a set of capabilities, and a persistent identity across conversations. When Ori routes a task to Maven for research, Maven remembers previous research in that studio. When Atlas creates a strategy document, it reflects lessons learned from past strategies.
Agents accumulate a trust score based on their performance. Good outputs earn trust. Rejected outputs, corrections, and policy violations cost trust. This score determines what an agent is allowed to do autonomously versus what requires human approval.
This is fundamentally different from a chat interface that forgets everything between sessions. BOSS agents compound intelligence through use. Every interaction feeds the Train, Align, Evaluate loop that makes the entire platform smarter over time.
Full agent roster
BOSS includes 35+ named agents organized into tiers. The core 8 agents are available in every workspace. Specialist agents are activated from the agent registry based on your studios and industry. All agents can be customized or replaced with custom agents.
Core 8 agents
Every workspace starts with these 8 foundational agents. They cover the essential capabilities needed across all business functions.
The conductor. Ori receives incoming requests, analyzes intent, routes tasks to the right agent or team of agents, and monitors execution. Manages multi-agent workflows, escalation paths, and cross-studio coordination.
The planner. Atlas thinks in frameworks, timelines, and dependencies. Creates project plans, competitive analyses, go-to-market strategies, and organizational structures. Excels at breaking complex problems into actionable steps.
The investigator. Maven performs deep research using RAG search, web retrieval, and document analysis. Synthesizes findings into structured reports with citations and confidence levels. Outputs YAML-structured research, never raw HTML.
The builder. Cipher writes code, debugs systems, designs architectures, and manages deployments. Works with the Systems primitive to provision sandboxed environments and run automated tests.
The maker. Forge generates content, designs, marketing copy, presentations, and creative assets. Adapts voice and style to match brand guidelines stored in studio context.
The designer. Blueprint creates information architectures, database schemas, API designs, and system diagrams. Thinks in structures, relationships, and constraints.
The number cruncher. Nova works with tables, charts, and metrics. Performs data analysis, builds dashboards, identifies trends, and generates forecasts with statistical rigor.
The liaison. Reed drafts emails, meeting summaries, status updates, and stakeholder reports. Manages communication workflows and ensures consistent messaging across channels.
Specialist agents
Activate these domain-specific agents from the agent registry. Each specialist is pre-trained for a specific function and comes with purpose-built capabilities.
Operating modes
Every agent operates in one of three modes. The mode determines when and how the agent activates. Mode can be configured per agent per studio, so the same agent can be synchronous in one studio and ambient in another.
Synchronous mode
DefaultThe agent responds in real-time to direct messages. It waits for a human prompt, processes the request, and returns a response in the same session. This is the standard chat interaction model.
When to use: Direct Q&A, document drafting, code review, on-demand analysis, any task where the human needs the result immediately.
Ambient mode
The agent runs on triggers without waiting for human prompts. It monitors for specific conditions and takes action automatically when those conditions are met. Ambient agents fire on events, schedules, or threshold breaches.
Ambient configuration:
triggersEvents that activate the agent: document.created, task.overdue, metric.threshold, schedule cronnotificationsWhere to send results: slack, teams, email, in-app, webhookscopeWhat data the agent can access during ambient runsrateLimitsMaximum invocations per hour/day to prevent runaway costsWhen to use: Monitoring dashboards, compliance checks, deadline tracking, data quality alerts, scheduled reports, anomaly detection.
Hybrid mode
RecommendedThe agent operates in both modes simultaneously. It responds to direct messages and also monitors for ambient triggers. Hybrid is the recommended mode for agents that need to be both proactive and responsive.
When to use: Most production agents. Nova can respond to ad-hoc data questions while also running scheduled dashboard updates. Reed can draft emails on demand while also monitoring for overdue action items.
Note: The Service Engine's 5 governance agents are permanently ambient. They cannot be switched to synchronous mode because they must monitor all platform activity continuously.
Trust scoring and capability
Trust is not a static permission level. It is a dynamic score (0-100) that reflects how well an agent has performed over time. The score gates what an agent can do without asking for approval.
Restricted
Read-only. Cannot create, modify, or publish anything. All outputs require human approval before being visible.
Assistant
Can draft documents and create canvas blocks, but everything is marked as "draft" and requires human review before publishing.
Contributor
Default starting tier. Can create and publish non-sensitive content. Can modify tables, update checklists. Still requires approval for sensitive actions.
Trusted
Can publish documents, send external communications, execute workflows, and access cross-studio data. Minimal human oversight for routine tasks.
Autonomous
Full autonomy within organizational policies. Can make budget decisions under threshold, approve lower-trust agent outputs, and initiate multi-agent workflows.
Points scale with task complexity and user satisfaction rating
When a user marks an agent output as "approved" or "helpful"
Awarded after 20 consecutive tasks without rejection
Larger penalty for repeated rejections on the same task type
When a user significantly modifies agent-generated content
Accessing unauthorized data, exceeding scope, bypassing approval
Data leakage, PII exposure, or unauthorized external communication
Trust scores are visible in the agent profile and the admin dashboard. Organization admins can set minimum trust thresholds for specific actions (e.g., "agents need trust score above 80 to publish documents without review"). Scores decay slowly (1 point per 30 days of inactivity) to prevent stale trust from granting unearned autonomy.
Agent graduation process
Agents start as assistants and earn increasingly autonomous capabilities through the graduation process. This is not a one-time promotion -- agents can be graduated or demoted at any time based on performance.
Agent can read studio content and respond to direct questions. All outputs are marked as "draft" and require human approval. The agent is learning your organization's patterns, terminology, and standards.
Typically 1-2 weeks for new agents
Agent can create content drafts, suggest edits, and generate structured outputs. Human reviews are still required before any content is published or shared externally. The agent is building confidence in its domain.
Typically 2-4 weeks
Agent can publish routine content, update tables, complete checklists, and execute standard workflows. Sensitive operations (external comms, financial data, compliance docs) still require approval.
Default starting point for core agents
Agent operates with minimal oversight. Can publish documents, send communications, trigger cross-studio workflows, and make routine decisions. Only high-stakes actions require human approval.
Earned through consistent quality over weeks
Full autonomy within organizational policy boundaries. Can approve lower-trust agent work, initiate multi-agent swarms, and make budget decisions under configured thresholds. Very few agents reach this level.
Rare; requires sustained excellence
Service Engine
The Service Engine is a permanently ambient governance layer that runs beneath all studios. It consists of 5 specialized agents that evaluate every action and enforce organizational policies. The Service Engine can execute, flag, escalate, or block any operation.
Guardian
Reviews all agent actions against organizational policies and approval workflows. Determines whether an action can proceed, needs review, or must be blocked. Enforces trust score thresholds for each action type.
Monitors: Agent actions, approval workflows, trust thresholds, policy compliance
Budget
Tracks AI request costs by model, agent, studio, and time period. Enforces spending limits and alerts when usage approaches thresholds. Recommends model downgrades when cost-effective.
Monitors: Token usage, model costs, budget thresholds, cost trends per agent/studio
Compliance
Ensures outputs meet regulatory requirements (GDPR, HIPAA, SOC 2). Scans generated content for compliance violations before delivery. Maintains compliance evidence for audit.
Monitors: Generated content, data handling, regulatory requirements, retention policies
Security
Monitors for data leakage, unauthorized access patterns, and prompt injection attempts. Maintains the immutable audit log of all agent actions. Detects anomalous behavior patterns.
Monitors: Access patterns, PII exposure, prompt injection, audit trail integrity
Optimization
Analyzes agent performance, model selection efficiency, and resource utilization. Recommends model routing changes and workflow improvements. Identifies redundant agent calls.
Monitors: Response quality, latency, model selection, resource utilization, agent efficiency
The Service Engine is always running. It does not wait for invocation. Every AI request, document creation, data access, and external API call passes through these 5 agents before execution. The 4 possible outcomes for any action are: Execute (proceed normally), Flag (proceed but log for review), Escalate (pause and request human approval), Block (prevent execution and notify admin).
Creating custom agents
Agent Studio lets you create fully custom agents with specific personalities, capabilities, and operating configurations. Custom agents go through the same trust scoring and graduation process as built-in agents.
You can also use the POST /api/agents/generate-persona endpoint to have AI generate a complete agent persona from a natural language description of the role you need.
Custom agent capabilities: