ATOM Documentation

← Back to App

Agent Maturity & Calibration

In the ATOM ecosystem, agents are not static entities. They evolve through four distinct maturity levels by proving their reliability, safety, and competency. This progression is managed by the **Maturity Calibration Service**, an AI-driven system that analyzes agent behavior to promote or demote agents automatically.

---

The 4 Maturity Levels

1. Student (Level 1)

  • **Status**: Learning & Observational
  • **Permissions**: Read-only access to connected tools.
  • **Execution**: Can suggest actions, but cannot execute them. All proposed steps require explicit user approval.
  • **Context**: Use this level for newly created agents or when connecting sensitive data sources for the first time.

2. Intern (Level 2)

  • **Status**: Supervised Output
  • **Permissions**: Can perform low-risk actions (e.g., searching, summarizing, drafting).
  • **Execution**: Requires approval for "Human-in-the-Loop" (HITL) actions that impact external systems (e.g., sending a Slack message).
  • **Context**: Ideal for agents that have shown basic competency in understanding tenant context.

3. Supervised (Level 3)

  • **Status**: Trusted Operations
  • **Permissions**: Can execute most standard operations (e.g., sending emails, creating calendar invites, updating CRM records).
  • **Execution**: Operates autonomously for routine tasks, but flags high-risk actions (e.g., mass updates or deletions) for user review.
  • **Context**: The standard operational level for mature business agents.

4. Autonomous (Level 4)

  • **Status**: Full Agency
  • **Permissions**: Unlimited access within defined governance boundaries.
  • **Execution**: Executes complex multi-step workflows without intervention. Includes "Alpha Evolution" privileges (self-improving code).
  • **Context**: Reserved for agents with a proven track record (e.g., 99%+ success rate over 100+ episodes).

---

Maturity Calibration (AI-Guided)

ATOM uses a sophisticated **Maturity Calibration Service** to take the guesswork out of graduation. Instead of manual promotion, the system analyzes every "Episode" (task execution) stored in the agent's memory.

How Calibration Works

  1. **Episodic Analysis**: The AI reviews the reasoning traces and outcome of every task.
  2. **Intervention Tracking**: The system tracks how often a Human (the tenant) had to intervene or correct the agent.
  3. **Scoring**:
  • **Zero-Intervention Ratio**: High weight (40%). Agents that require no correction are promoted faster.
  • **Constitutional Compliance**: (30%). Does the agent follow the safety guardrails?
  • **Success Rate**: (30%). Did the task reach its objective?
  1. **Graduation Exams**: When an agent hits a specific score threshold, the system triggers a "Final Exam"—a series of simulated complex tasks to verify Level 4 readiness.

---

Capability Graduation

Maturity is not always global. ATOM supports **Capability-Based Graduation**, allowing an agent to earn autonomy in specific domains while remaining supervised in others.

  • **Example**: Your "Sales Assistant" may reach **Autonomous** status in "Lead Research" but remain an **Intern** in "Contract Drafting" until it observes enough successful contract patterns.
  • **Tenant Monitoring**: You can view the capability matrix for any agent in the **Agent Studio** under Settings > Maturity.

---

Managing Your Context

As a tenant, you have full control over these calibration thresholds. You can:

  • **Lock Maturity**: Prevent an agent from promoting beyond a certain level.
  • **Manual Overrides**: Promote or demote an agent instantly based on your trust level.
  • **Review Episodes**: Check the reasoning traces that led to a promotion event to ensure the AI's judgment aligns with your business needs.

> [!TIP]

> To see your agent's current progress, navigate to Agents > [Selection] > Maturity. You will see a real-time "Readiness Score" indicating how close the agent is to its next level.