Agent Reasoning¶

Agent reasoning enables an AI Agent to interpret goals, evaluate available information, select appropriate actions, and adapt its behavior based on observations and outcomes.

📖 Overview¶

Planning determines what an agent should do.

Reasoning helps determine why a particular action should be taken and what should happen next.

A production AI Agent continuously operates across a decision loop:

Goal
 ↓
Understand Context
 ↓
Assess Situation
 ↓
Consider Available Actions
 ↓
Select Action
 ↓
Execute
 ↓
Observe Result
 ↓
Evaluate
 ↓
Decide Next Step

Reasoning becomes particularly important when the agent must:

• Interpret ambiguous requests
• Select between multiple possible actions
• Determine which tool to use
• Evaluate tool results
• Handle incomplete information
• Identify errors
• Decide whether additional information is required
• Adapt to changing conditions
• Determine whether the task has been completed
• Decide whether to retry, re-plan, or escalate

In enterprise systems, reasoning should not mean unrestricted autonomy.

Instead, the objective is:

Use reasoning within controlled execution boundaries, policies, tools, budgets, and validation mechanisms.

🎯 Learning Objectives¶

After completing this chapter, you will understand:

• What reasoning means in AI Agents
• The relationship between reasoning and planning
• Reasoning in the agent execution loop
• Action selection
• Observation-driven reasoning
• Tool-aware reasoning
• ReAct-style agent execution
• Reflection-based reasoning
• Self-correction
• Structured reasoning approaches
• Reasoning with constraints
• Reasoning under uncertainty
• Reasoning failures
• Reasoning evaluation
• Reasoning observability
• Production reasoning guardrails
• How reasoning fits into enterprise AI Agent architecture

1. What Is Agent Reasoning?¶

Agent reasoning is the process through which an agent interprets its current state, evaluates available options, and determines the next appropriate action.

A simplified model is:

Current State
     ↓
Understand Situation
     ↓
Evaluate Options
     ↓
Select Action
     ↓
Execute
     ↓
Observe
     ↓
Update State
     ↓
Reason Again

For example:

"Find the cause of a failed payment."

The agent may need to determine:

What payment?
     ↓
Retrieve transaction
     ↓
What is the current status?
     ↓
Inspect payment result
     ↓
What caused the failure?
     ↓
Check error information
     ↓
What action is allowed?
     ↓
Apply appropriate resolution

The important characteristic is that the next action depends on information available at runtime.

2. Reasoning vs Planning¶

Planning and reasoning are closely related.

However, they serve different purposes.

Planning¶

Planning determines:

What sequence of tasks should be performed?

Goal
 ↓
Task A
 ↓
Task B
 ↓
Task C

Reasoning¶

Reasoning determines:

What should the agent decide at a particular point in execution?

Current State
 ↓
Evaluate Situation
 ↓
Choose Action

Together:

                Goal
                  ↓
              Reasoning
                  ↓
                Plan
                  ↓
              Execute
                  ↓
              Observe
                  ↓
              Reasoning
                  ↓
          Continue / Re-plan

Planning provides structure.

Reasoning provides decision-making within that structure.

3. Reasoning in the Agent Loop¶

A typical agent execution loop can be represented as:

┌───────────────┐
│     Goal      │
└───────┬───────┘
        ↓
┌───────────────┐
│    Reason     │
└───────┬───────┘
        ↓
┌───────────────┐
│     Plan      │
└───────┬───────┘
        ↓
┌───────────────┐
│  Select Tool  │
└───────┬───────┘
        ↓
┌───────────────┐
│ Execute Tool  │
└───────┬───────┘
        ↓
┌───────────────┐
│    Observe    │
└───────┬───────┘
        ↓
┌───────────────┐
│ Evaluate      │
│ Result        │
└───────┬───────┘
        ↓
   ┌────┴─────┐
   ↓          ↓
Complete    Continue
              │
              └──────→ Reason

Reasoning can therefore occur repeatedly throughout a task rather than only once.

4. State-Based Reasoning¶

An agent should reason based on its current state.

A conceptual agent state may include:

Agent State
├── User Goal
├── Current Task
├── Conversation Context
├── Memory
├── Available Tools
├── Tool Results
├── Previous Actions
├── Constraints
├── Policies
└── Execution Status

The reasoning process evaluates this state to determine the next action.

Agent State
     ↓
Reasoning
     ↓
Next Action
     ↓
State Change
     ↓
Reasoning

This is why state management is a fundamental part of production agent architecture.

5. Action Selection¶

An agent may have multiple possible actions.

For example:

User:
"Why is my order delayed?"

The agent may have access to:

Order Lookup Tool
Shipment Tracking Tool
Customer Database
Knowledge Base
Support Ticket API

The agent must determine which action is appropriate.

User Request
     ↓
Understand Intent
     ↓
Available Capabilities
     ↓
Evaluate Options
     ↓
Select Tool
     ↓
Execute

Good action selection should consider:

• Tool capability
• Required inputs
• Tool availability
• User intent
• Current state
• Permissions
• Business rules
• Cost
• Latency
• Risk

6. Tool-Aware Reasoning¶

Reasoning becomes more useful when the agent understands the capabilities and constraints of its tools.

Consider:

Goal:
Check account balance

Available tools:

get_customer_profile()
get_account_balance()
create_payment()

A suitable reasoning process should identify:

Required information:
Account balance

Relevant capability:
get_account_balance()

Irrelevant capabilities:
create_payment()

The agent should therefore avoid unnecessary actions.

Conceptually:

Goal
 ↓
Capability Matching
 ↓
Tool Selection
 ↓
Tool Execution
 ↓
Observation

Tool-aware reasoning is particularly important because unnecessary tool calls increase:

• Latency
• Cost
• Failure probability
• Security exposure

7. Observation-Driven Reasoning¶

An agent often cannot determine the correct next step until a tool produces a result.

Example:

Check Payment
      ↓
    Result
      ↓
 ┌────┴───────┐
 ↓            ↓
Successful   Failed
 ↓            ↓
Complete    Investigate

The agent therefore reasons from observations.

Action
 ↓
Observation
 ↓
Interpretation
 ↓
Decision
 ↓
Next Action

This creates a feedback loop:

Reason
  ↓
Act
  ↓
Observe
  ↓
Reason
  ↓
Act
  ↓
Observe

This loop is one of the defining characteristics of agentic execution.

8. ReAct-Style Reasoning¶

One widely used agent pattern is ReAct, which combines reasoning and action.

Conceptually:

Reason
 ↓
Act
 ↓
Observe
 ↓
Reason
 ↓
Act
 ↓
Observe

For example:

Goal:
Find the status of an order.

Reason
 ↓
Need order information

Act
 ↓
Call Order API

Observe
 ↓
Order is shipped

Reason
 ↓
Need shipment tracking information

Act
 ↓
Call Tracking API

Observe
 ↓
Shipment arrives tomorrow

Final Response

The key idea is that the agent uses observations from previous actions to determine what to do next.

Important: Production systems should capture useful decisions, actions, results, and state transitions without exposing or depending on private chain-of-thought.

9. Structured Reasoning¶

Reasoning does not need to rely entirely on free-form natural language.

Production agents can use structured decision representations.

For example:

{
  "decision": "retrieve_order",
  "reason": "order_id_required_for_status_lookup",
  "tool": "order_lookup",
  "next_state": "ORDER_LOOKUP"
}

Structured reasoning can make agent execution easier to:

• Validate
• Trace
• Test
• Monitor
• Audit
• Control

A production architecture can therefore separate:

LLM Decision
      ↓
Structured Action
      ↓
Policy Validation
      ↓
Tool Execution

10. Reasoning with Constraints¶

Enterprise agents operate under constraints.

Examples include:

Security Policies
Business Rules
Tool Permissions
Budget Limits
Time Limits
Compliance Requirements
Data Access Policies
Human Approval Requirements

Consider:

Goal:
Issue customer refund

The agent may determine:

Refund Amount = $500

Policy:

< $1,000 → Agent may proceed
≥ $1,000 → Human approval required

The reasoning process must therefore incorporate the policy:

Determine Action
       ↓
Check Policy
       ↓
 ┌─────┴─────┐
 ↓           ↓
Allowed    Approval
 ↓           ↓
Execute    Human Review

This creates bounded reasoning.

11. Reasoning Under Uncertainty¶

Real-world agent environments frequently contain incomplete or conflicting information.

For example:

Customer Request
       ↓
Order ID Missing
       ↓
Customer Account Available
       ↓
Multiple Orders Found

The agent must determine what to do next.

Possible actions:

Ask User
      OR
Search Additional Information
      OR
Request Clarification

A safe agent should not simply guess.

Instead:

Insufficient Information
        ↓
Assess Uncertainty
        ↓
Can It Be Resolved Safely?
        │
     ┌──┴──┐
     ↓     ↓
    Yes    No
     ↓     ↓
 Gather   Ask /
 Context  Escalate

A key production principle is:

When uncertainty materially affects the outcome, prefer clarification or escalation over unsupported assumptions.

12. Reasoning with Memory¶

Memory provides additional context for reasoning.

Current Goal
     ↓
Retrieve Relevant Memory
     ↓
Combine with Current State
     ↓
Reason
     ↓
Select Action

For example:

Previous Interaction:
Customer prefers email communication.

Current Request:
"Send me the update."

Agent Memory
      ↓
Preferred Channel = Email
      ↓
Reasoning
      ↓
Select Email Tool

This creates the relationship:

Memory
   ↓
Context
   ↓
Reasoning
   ↓
Action

Memory architecture is covered in:

Part VI — Agent Memory.

13. Reasoning and Reflection¶

Reasoning determines an action.

Reflection evaluates the outcome.

Reason
 ↓
Act
 ↓
Observe
 ↓
Reflect
 ↓
Evaluate

For example:

Agent decides:
"Use database query."

       ↓

Query executed.

       ↓

Result:
No matching records.

       ↓

Reflection:
"The current query did not produce sufficient information."

       ↓

Next decision:
Try an alternative search strategy.

This distinction is important:

Reasoning
= Decide what to do

Reflection
= Evaluate what happened

Reflection and self-correction are covered in:

10. Reflection & Self-Correction.

14. Reasoning and Re-Planning¶

Reasoning can trigger re-planning.

Initial Plan
     ↓
Execute
     ↓
Observe
     ↓
Unexpected Result
     ↓
Reason
     ↓
Re-plan
     ↓
Continue

For example:

Plan:
1. Retrieve database record
2. Analyze record
3. Generate report

Execution:
Database unavailable

Reasoning:
Database path cannot be completed.

Re-plan:
1. Query replica
2. If unavailable → use approved fallback
3. Continue analysis

Thus:

Reasoning
    ↓
Re-Planning
    ↓
Execution

15. Reasoning Strategies¶

Different agent architectures can use different reasoning strategies.

Common approaches include:

• Direct decision making
• ReAct-style reasoning
• Reflection
• Self-correction
• Planning-based reasoning
• Tool-assisted reasoning
• Structured decision making
• Retrieval-supported reasoning
• Rule-constrained reasoning

These approaches can be combined.

For example:

Planning
   +
Tool Calling
   +
Observation
   +
Reflection
   +
Re-planning

The correct strategy depends on:

• Task complexity
• Reliability requirements
• Latency requirements
• Cost constraints
• Risk level
• Tool environment

16. Direct Reasoning¶

For simple tasks, an agent may not require elaborate planning.

User
 ↓
Agent
 ↓
Decision
 ↓
Action

Example:

"What's the weather?"

The agent may simply:

Understand Request
 ↓
Select Weather Tool
 ↓
Call Tool
 ↓
Return Result

Adding unnecessary reasoning stages can increase:

• Latency
• Token consumption
• Cost
• Failure opportunities

Therefore:

Use the simplest reasoning strategy that reliably solves the task.

17. Multi-Step Reasoning¶

Complex tasks may require multiple decisions.

Goal
 ↓
Decision 1
 ↓
Action 1
 ↓
Observation
 ↓
Decision 2
 ↓
Action 2
 ↓
Observation
 ↓
Decision 3
 ↓
Final Result

For example:

Investigate failed transaction
       ↓
Retrieve transaction
       ↓
Check payment status
       ↓
Inspect error
       ↓
Check account state
       ↓
Determine resolution

Each observation influences the next decision.

18. Reasoning with Retrieval¶

Agents can use RAG as a reasoning support mechanism.

Agent
 ↓
Determine Information Needed
 ↓
Retriever
 ↓
Relevant Context
 ↓
Reason
 ↓
Action

For example:

Customer asks:
"Can I cancel this subscription?"

Agent
 ↓
Retrieve cancellation policy
 ↓
Check subscription state
 ↓
Reason
 ↓
Determine whether cancellation is allowed
 ↓
Respond / Execute

This creates:

RAG
 ↓
Grounded Context
 ↓
Agent Reasoning
 ↓
Action

The advanced autonomous retrieval patterns are explored in Part VII — Agentic AI & Multi-Agent Systems.

19. Reasoning and Policies¶

Reasoning should not bypass enterprise policy.

Consider:

Agent Decision
      ↓
Policy Engine
      ↓
 ┌────┴────┐
 ↓         ↓
Allowed   Denied
 ↓         ↓
Execute   Stop / Escalate

The LLM can propose an action.

The policy layer determines whether the action is permitted.

This creates an important architectural separation:

LLM
 ↓
Decision
 ↓
Policy
 ↓
Execution

This is safer than allowing the model to directly control unrestricted enterprise systems.

20. Reasoning and Risk¶

Not every agent action carries the same risk.

A useful conceptual model is:

Low Risk
   ↓
Read Information
   ↓
Generate Summary
   ↓
Draft Recommendation
   ↓
Modify Data
   ↓
Execute Transaction
   ↓
High Risk

Higher-risk actions should require stronger controls.

Low Risk
   ↓
Automatic Execution

Medium Risk
   ↓
Policy Validation

High Risk
   ↓
Human Approval

This allows enterprises to implement risk-based autonomy.

21. Reasoning Budgets¶

Reasoning has a cost.

An agent may perform too many iterations:

Reason
 ↓
Act
 ↓
Observe
 ↓
Reason
 ↓
Act
 ↓
Observe
 ↓
...

Production systems should therefore define limits such as:

Maximum Iterations
Maximum Tool Calls
Maximum Runtime
Maximum Token Budget
Maximum Cost
Maximum Planning Depth

Example:

Agent Run
 │
 ├── Iteration Limit = 10
 ├── Tool Call Limit = 20
 ├── Runtime Limit = 120 sec
 └── Cost Limit = $0.50

If a limit is reached:

Stop
 ↓
Return Safe Result
       OR
Escalate

22. Reasoning Failures¶

Agent reasoning can fail in several ways.

Incorrect Action Selection¶

User Goal
 ↓
Wrong Tool
 ↓
Incorrect Result

Insufficient Context¶

Incomplete Context
 ↓
Incorrect Decision

Invalid Assumption¶

Missing Information
 ↓
Agent Assumes Value
 ↓
Incorrect Action

Tool Misinterpretation¶

Tool Result
 ↓
Incorrect Interpretation
 ↓
Incorrect Next Action

Reasoning Loop¶

Reason
 ↓
Action
 ↓
Observation
 ↓
Reason
 ↓
Same Action
 ↓
...

Over-Reasoning¶

Simple Task
 ↓
Unnecessary Decisions
 ↓
Higher Cost
 ↓
Higher Latency

Policy Violation¶

Agent Decision
 ↓
Restricted Action
 ↓
Policy Violation

This is why reasoning must operate inside security and policy boundaries.

23. Reasoning Failure Handling¶

A production agent should detect reasoning failures.

Decision
 ↓
Validation
 ↓
 ┌──────────┴──────────┐
 ↓                     ↓
Valid                 Invalid
 ↓                     ↓
Execute             Correct / Re-plan
                       ↓
                    Retry
                       ↓
                   Escalate

Possible recovery strategies include:

• Retry
• Re-plan
• Request more information
• Use another tool
• Fall back to deterministic logic
• Escalate to a human
• Terminate safely

24. Reasoning and Deterministic Logic¶

Not every decision should be delegated to an LLM.

For deterministic business rules:

Business Rule
      ↓
Deterministic Code

For ambiguous interpretation:

Natural Language
      ↓
LLM Reasoning

A strong enterprise architecture combines both:

             Agent
               │
       ┌───────┴────────┐
       ↓                ↓
LLM Reasoning      Deterministic Rules
       │                │
       └───────┬────────┘
               ↓
          Validated Action

This improves:

• Predictability
• Reliability
• Compliance
• Testability

25. Structured Decision Layer¶

A production agent can use an intermediate decision layer.

User Request
     ↓
LLM
     ↓
Structured Decision
     ↓
Policy Validation
     ↓
Action Router
     ↓
Tool

For example:

{
  "action": "lookup_order",
  "arguments": {
    "order_id": "ORD-12345"
  }
}

The execution layer can validate:

• Action name
• Arguments
• Permissions
• Input schema
• Policy
• Risk level

before executing the action.

26. Reasoning Observability¶

Reasoning itself should be observable through decision metadata and execution traces, rather than relying on exposing private chain-of-thought.

Useful telemetry includes:

Agent Run
 │
 ├── Goal
 ├── Current State
 ├── Selected Action
 ├── Tool
 ├── Tool Arguments
 ├── Tool Result
 ├── Decision Outcome
 ├── Retry
 ├── Re-plan
 ├── Policy Result
 └── Final Outcome

This allows engineers to answer:

• Which tool was selected?
• How many tool calls occurred?
• Where did the agent fail?
• Did the agent re-plan?
• How many iterations were required?
• Which policy blocked an action?
• How much did the execution cost?

This is covered further in:

Part VI — Agent Observability.

27. Evaluating Reasoning¶

Reasoning quality should not be evaluated only by inspecting the final response.

Important evaluation dimensions include:

Reasoning Quality
├── Correct Action Selection
├── Tool Selection
├── Tool Argument Accuracy
├── Goal Completion
├── Error Recovery
├── Re-Planning Quality
├── Policy Compliance
├── Efficiency
├── Latency
└── Cost

For example:

Task
 ↓
Agent Decision
 ↓
Tool
 ↓
Result
 ↓
Final Outcome

Evaluation can examine the complete execution trajectory.

The dedicated agent evaluation topic is covered later in the Part VI observability section.

28. Reasoning Efficiency¶

A good reasoning strategy should minimize unnecessary work.

Consider two approaches.

Inefficient¶

User
 ↓
Reason
 ↓
Reason
 ↓
Reason
 ↓
Tool
 ↓
Reason
 ↓
Reason
 ↓
Answer

Efficient¶

User
 ↓
Understand
 ↓
Select Tool
 ↓
Execute
 ↓
Observe
 ↓
Answer

The objective is not to maximize reasoning steps.

The objective is:

Use enough reasoning to reliably achieve the task, but no more than necessary.

29. Enterprise Reasoning Architecture¶

A production-oriented architecture can separate reasoning, policy, and execution.

                         User
                           │
                           ▼
                     Agent Gateway
                           │
                           ▼
                      Agent State
                           │
                           ▼
                       Reasoner
                           │
                           ▼
                    Proposed Action
                           │
                           ▼
                    Policy Engine
                           │
                    ┌──────┴──────┐
                    ↓             ↓
                 Allowed        Denied
                    ↓             ↓
                Tool Router    Escalation
                    │
          ┌─────────┼─────────┐
          ↓         ↓         ↓
        Tool A    Tool B    Tool C
          │         │         │
          └─────────┼─────────┘
                    ↓
                 Result
                    │
                    ▼
                 Evaluate
                    │
              ┌─────┴─────┐
              ↓           ↓
          Complete      Continue
                           │
                           ▼
                       Reason Again

This architecture provides:

Reasoning
   +
Policy
   +
Execution
   +
Observation

rather than allowing unrestricted model-driven execution.

30. Reasoning in a Production Agent Lifecycle¶

A complete agent lifecycle can be represented as:

                    Goal
                     │
                     ▼
                  Context
                     │
                     ▼
                  Reason
                     │
                     ▼
                   Plan
                     │
                     ▼
               Select Action
                     │
                     ▼
              Policy Validation
                     │
                     ▼
                Execute Tool
                     │
                     ▼
                 Observe
                     │
                     ▼
                 Evaluate
                     │
              ┌──────┴──────┐
              ↓             ↓
          Complete        Continue
                            │
                            ▼
                         Reflect
                            │
                            ▼
                       Re-plan
                            │
                            └──────→ Reason

This forms the core execution loop of an AI Agent.

31. Practical Example — Customer Support Agent¶

Consider:

"My payment failed. Please check what happened."

The agent may reason through:

Understand Request
       ↓
Need Customer / Transaction Information
       ↓
Retrieve Transaction
       ↓
Observe Result
       ↓
Payment Failed
       ↓
Retrieve Failure Details
       ↓
Observe Error
       ↓
Determine Resolution
       ↓
Check Policy
       ↓
Execute Allowed Action
       ↓
Verify Result
       ↓
Respond

The important point is that each observation influences the next decision.

32. Practical Example — Software Engineering Agent¶

Consider:

"Investigate why the build is failing."

The agent might perform:

Understand Task
      ↓
Inspect Build Result
      ↓
Identify Failure
      ↓
Inspect Relevant Code
      ↓
Determine Likely Cause
      ↓
Inspect Dependencies
      ↓
Propose Fix
      ↓
Apply Approved Change
      ↓
Run Tests
      ↓
Observe
      ↓
 ┌────┴────┐
 ↓         ↓
Pass      Fail
 ↓         ↓
Complete  Re-plan

The agent does not need to execute every possible diagnostic tool.

Reasoning should determine which investigation step provides the most useful next information.

33. Reasoning and Agent Autonomy¶

Reasoning is one of the mechanisms that enables autonomy.

However:

Reasoning ≠ Unlimited Autonomy

A production agent should operate within:

Tools
Policies
Permissions
Budgets
Guardrails
Human Approval
Execution Limits

Therefore:

Autonomy
    =
Reasoning
    +
Actions
    +
Feedback
    +
Boundaries

This is a critical enterprise AI engineering principle.

34. Reasoning vs Multi-Agent Reasoning¶

This chapter focuses on reasoning within an individual AI Agent.

Part VI
Single Agent
     ↓
Reason
     ↓
Plan
     ↓
Act
     ↓
Observe

Part VII extends reasoning into:

Agent A
   ↓
Agent B
   ↓
Agent C
   ↓
Coordination
   ↓
Collective Decision

Therefore, concepts such as:

• Multi-agent reasoning
• Agent debate
• Supervisor reasoning
• Swarm coordination
• Collaborative reasoning

belong to Part VII — Agentic AI & Multi-Agent Systems.

35. Key Engineering Principles¶

1. Reasoning Should Serve the Goal¶

Do not add reasoning complexity without a practical purpose.

2. Reason from Current State¶

Agent decisions should incorporate the latest available information.

3. Use Tools When External Information Is Required¶

The model should not invent information that can be retrieved from authoritative systems.

4. Validate Actions¶

LLM-generated decisions should pass through appropriate validation and policy controls.

5. Separate Decision from Execution¶

The model can propose an action while deterministic infrastructure controls execution.

6. Handle Uncertainty Explicitly¶

When information is insufficient, ask, retrieve, or escalate rather than guessing.

7. Use Reflection for Recovery¶

Evaluate outcomes and determine whether corrective action is necessary.

8. Control Reasoning Loops¶

Set iteration, time, token, tool, and cost limits.

9. Observe Decisions¶

Capture structured execution telemetry without depending on private chain-of-thought.

10. Prefer Controlled Autonomy¶

Enterprise agents should operate within clearly defined boundaries.

36. Key Takeaways¶

• Reasoning enables AI Agents to make decisions during task execution.
• Planning defines the sequence of tasks; reasoning determines decisions within that execution.
• Agent reasoning is often iterative and observation-driven.
• Tool-aware reasoning helps agents select appropriate capabilities.
• ReAct-style execution combines reasoning, action, and observation.
• Structured decisions make agent behavior easier to validate and operate.
• Reasoning should account for policies, permissions, constraints, and risk.
• Agents should explicitly handle uncertainty rather than blindly guessing.
• Reflection evaluates outcomes and can trigger self-correction or re-planning.
• Deterministic business rules should remain deterministic where appropriate.
• Production systems should control reasoning depth, iterations, latency, and cost.
• Reasoning should be observable through structured decision and execution telemetry.
• Reasoning quality should be evaluated using task outcomes and execution trajectories.
• The goal is not maximum reasoning; it is reliable and efficient decision-making.
• Enterprise AI Agents should use controlled autonomy rather than unrestricted autonomy.

🔗 Related Topics¶

Previous¶

08. Planning & Task Decomposition

Next¶

10. Reflection & Self-Correction

Related¶

• AI Agent Fundamentals
• Tool Calling & Function Calling
• Building & Orchestrating Tools
• Agent Memory
• Agent Observability
• Agent Security

Enterprise AI Engineering Handbook
Building Production-Grade Enterprise AI Systems — One Chapter at a Time.