The TDAO Loop

Every AiQarus agent executes using the Think-Decide-Act-Observe (TDAO) loop—a structured reasoning framework that ensures transparency and traceability.

Overview


    ┌─────────────────────────────────────────────────────────┐
    │                                                         │
    │   ┌─────────┐     ┌─────────┐     ┌─────────┐     ┌─────────┐
    │   │  THINK  │────▶│ DECIDE  │────▶│   ACT   │────▶│ OBSERVE │
    │   └─────────┘     └─────────┘     └─────────┘     └─────────┘
    │        │               │               │               │
    │        ▼               ▼               ▼               ▼
    │   [Analyze]       [Choose]        [Execute]      [Summarize]
    │                       │
    │                       ├── Complete ──▶ Exit with Result
    │                       ├── Continue ──▶ Loop ─────────────┘
    │                       └── Fail ──────▶ Exit with Error
    │
    └─────────────────────────────────────────────────────────┘

The Four Phases

1. THINK

The agent analyzes the current situation.

Input:

Original user input/goal
Previous observations (if any)
Current context and state

Output:

Structured reasoning trace
Identification of what’s known vs unknown
Assessment of progress toward goal

Example Think Output:


{
  "phase": "think",
  "reasoning": "The user wants to offboard Sarah Chen. I have gathered her employee ID and access information. I've identified that she has AWS admin access which is high-risk given she's leaving for a competitor. I need to prioritize revoking this access before proceeding with other offboarding tasks.",
  "knowns": [
    "Employee: Sarah Chen",
    "Department: Engineering",
    "AWS access level: Admin",
    "Departure type: Competitor"
  ],
  "unknowns": [
    "Pending code reviews",
    "Active customer relationships"
  ],
  "progress": "40% - Information gathered, ready for action planning"
}

2. DECIDE

The agent chooses what to do next.

Possible Decisions:

Decision	Meaning	Next Step
Continue	More work to do	Proceed to ACT with specific action
Complete	Goal achieved	Exit with final output
Fail	Cannot proceed	Exit with error explanation

Example Decide Output:


{
  "phase": "decide",
  "decision": "continue",
  "action": {
    "tool": "okta_revoke_access",
    "parameters": {
      "user_id": "sarah.chen@company.com",
      "application": "aws-production"
    },
    "justification": "Revoking AWS admin access first due to high risk (competitor departure + production access)"
  },
  "risk_score": 85,
  "requires_approval": true
}

3. ACT

The agent executes the decided action.

What Happens:

Tool is invoked with specified parameters
External systems are called
Results are captured
All activity is traced

Example Act Output:


{
  "phase": "act",
  "tool": "okta_revoke_access",
  "input": {
    "user_id": "sarah.chen@company.com",
    "application": "aws-production"
  },
  "output": {
    "success": true,
    "revoked_at": "2026-01-15T14:30:00Z",
    "affected_resources": 15
  },
  "duration_ms": 1250
}

4. OBSERVE

The agent summarizes results to inform the next iteration.

Example Observe Output:


{
  "phase": "observe",
  "summary": "Successfully revoked AWS production admin access for Sarah Chen. 15 resources were affected. The user can no longer access production databases, EC2 instances, or S3 buckets.",
  "state_changes": [
    "AWS admin access: active → revoked"
  ],
  "next_considerations": [
    "Proceed with GitHub access revocation",
    "Document which resources were affected for audit"
  ]
}

Step Records

Each TDAO iteration creates a Step record:


{
  "step_id": "step_456",
  "run_id": "run_123",
  "step_number": 3,
  "phases": {
    "think": { ... },
    "decide": { ... },
    "act": { ... },
    "observe": { ... }
  },
  "tokens_used": {
    "input": 2500,
    "output": 800
  },
  "duration_ms": 4500,
  "started_at": "2026-01-15T14:29:55Z",
  "completed_at": "2026-01-15T14:30:00Z"
}

Why TDAO Matters

1. Transparency

Every decision is recorded with full reasoning:

Why the agent chose this action
What alternatives were considered
How risk was assessed

2. Debuggability

When something goes wrong, you can trace:

Which step failed
What the agent was thinking
What information it had

3. Auditability

Regulators and compliance teams can review:

The complete decision-making process
Justifications for each action
Risk assessments and approvals

4. Reproducibility

Given the same input and state:

The same reasoning should occur
Behavior is predictable
Testing is reliable

Viewing TDAO in the UI

The Run Detail view shows each step:


┌─────────────────────────────────────────────────────────────────────┐
│  Run: Off-boarding Sarah Chen                                        │
│  Status: Completed                                                   │
│                                                                      │
│  ────────────────────────────────────────────────────────────────── │
│                                                                      │
│  Step 1: Gather Information                          ✓ 2.1s         │
│  ├─ THINK: Analyzing offboarding request...                         │
│  ├─ DECIDE: Continue → fetch_employee_data                          │
│  ├─ ACT: Called okta_get_user → success                             │
│  └─ OBSERVE: Retrieved employee profile and access list             │
│                                                                      │
│  Step 2: Assess Risk                                 ✓ 1.8s         │
│  ├─ THINK: Evaluating access scope and departure type...            │
│  ├─ DECIDE: Continue → calculate risk score                         │
│  ├─ ACT: Risk assessment complete → score: 85/100                   │
│  └─ OBSERVE: High risk due to competitor + admin access             │
│                                                                      │
│  Step 3: Revoke AWS Access                           ✓ 3.5s         │
│  ├─ THINK: Prioritizing AWS due to high risk...                     │
│  ├─ DECIDE: Continue → revoke_aws_access (approval required)        │
│  │   └─ 🛑 PAUSED for approval (approved by john@company.com)       │
│  ├─ ACT: Called okta_revoke_access → success                        │
│  └─ OBSERVE: AWS access revoked, 15 resources affected              │
│                                                                      │
│  [Expand All]  [Verify Audit Chain]  [Export]                       │
│                                                                      │
└─────────────────────────────────────────────────────────────────────┘

Configuration

Custom Phase Prompts

Override default prompts for each phase:


phase_prompts:
  think: "Carefully analyze the current situation. Consider all available information and identify what you know vs what you need to find out."
  decide: "Based on your analysis, decide the next action. Always explain your reasoning."
  observe: "Summarize what happened and what it means for the overall goal."

Phase Timeouts

Set timeouts for individual phases:


phase_timeouts:
  think: 30s
  decide: 10s
  act: 60s
  observe: 15s