Learning Objectives

After completing this lecture, you will be able to:

• Understand the key concepts of emerging ai capabilities: agents, reasoning, and tools in a government context
• Connect emerging ai capabilities: agents, reasoning, and tools to your agency's AI initiatives
• Identify next steps for applying these concepts in your role

Key Topics Covered

-
Agentic AI, chain-of-thought reasoning, tool use, multi-step workflows

-
What's coming and what it means for government

-
Government context for emerging ai capabilities: agents, reasoning, and tools

-
Practical applications and next steps

Why This Matters for Government

Government agencies face unique challenges when it comes to AI adoption. This lecture addresses these challenges head-on by providing analysts, project leads, team supervisors with the knowledge and frameworks needed to navigate AI in the public sector responsibly and effectively.

As part of the L2 (AI Practitioner) curriculum, this lecture builds on the foundational principle that every AI system in government ultimately serves citizens. Whether you are working with AI tools daily or setting strategy for your agency, understanding emerging ai capabilities: agents, reasoning, and tools is essential for responsible, effective government AI adoption.

======================================================================

TRANSCRIPT: Emerging AI Capabilities: Agents, Reasoning, and Tools

======================================================================

What you will learn: What agentic AI is and how it differs from traditional chatbots; reasoning capabilities and chain-of-thought; tool use and how AI can call functions or access external systems; multi-step workflows and planning; risks specific to agentic AI systems; government applications.

We're at the frontier of AI capabilities. Beyond language models that respond to prompts, we're seeing AI systems that can act autonomously, reason through multi-step problems, use tools to access external systems, and plan complex workflows.

This lecture covers these emerging capabilities and their implications for government.

WHY THIS MATTERS FOR GOVERNMENT

Agentic AI could transform how government operates. Instead of an AI system that responds to queries, imagine a system that autonomously identifies cases that need follow-up, researches relevant regulations, drafts correspondence, and schedules meetings. These are no longer distant science fiction -- they're technically feasible and some agencies are experimenting with them.

But agentic AI also introduces new risks. Autonomous systems can make errors at scale. Systems that use tools to access external systems might access the wrong information or cause unintended consequences.

Understanding these capabilities and their risks is essential for responsible government adoption.

AGENTIC AI VERSUS REACTIVE CHATBOTS

A traditional LLM (chatbot) is reactive. You send a prompt, it responds. It doesn't do anything beyond generating text.

Agentic AI is proactive. It's given a goal and autonomously takes actions to achieve that goal. It might:

-- Break a complex goal into sub-goals

-- Plan steps to achieve those sub-goals

-- Execute steps

-- Observe outcomes

-- Adjust plans based on outcomes

-- Repeat until the goal is achieved

Example: "Process benefits applications that have been pending for more than 30 days."

An agentic system might:

• Query the database for applications pending >30 days
• For each application, retrieve the file
• Analyze the application for completeness
• If incomplete, generate a request letter asking for missing information
• If complete, check eligibility against regulations
• Generate a decision letter
• Route the letter to a supervisor for review
• Log what it did

This is fundamentally different from a chatbot that responds to "What should I do about this pending application?"

REASONING AND CHAIN-OF-THOUGHT

Traditional LLMs sometimes jump to conclusions without showing reasoning. Advanced AI systems can show chain-of-thought reasoning: thinking through a problem step-by-step, showing their work.

Example: "Is an applicant earning $50,000 annually eligible for benefits if the limit is 300% of federal poverty level? For a family of 3, 300% of FPL is $75,000. So yes, the applicant is eligible."

Chain-of-thought reasoning makes the system's thinking visible. This is valuable for government because:

-- You can check if the reasoning is correct

-- You can understand how the system reached a conclusion

-- If the conclusion is wrong, you can identify where the reasoning broke down

TOOL USE AND EXTERNAL INTEGRATION

An agentic system can be given tools: functions it can call to access external systems.

Tools might include:

-- Database queries (retrieve applicant information, check records)

-- Document retrieval (find regulations, retrieve forms)

-- Calculations (compute benefit amounts, check eligibility)

-- Communications (send emails, post notifications)

Example: A system trying to determine if an applicant is eligible for a program might:

• Call a "retrieve_regulations" tool to get the current eligibility requirements
• Call a "query_applicant_data" tool to get the applicant's information
• Use reasoning to compare the applicant's information to requirements
• Call a "generate_decision_letter" tool to create the decision

The advantage: the system is grounded in actual regulations and applicant data, not relying on its training data which might be outdated.

The risk: if the tools are called incorrectly or return unexpected information, the system might make wrong decisions.

MULTI-STEP WORKFLOWS AND PLANNING

Overview

Agentic systems can plan complex workflows with multiple steps, dependencies, and decision points.

Example workflow for processing a complex permit:

• Retrieve permit application
• Check if complete; if not, request missing information
• Retrieve relevant environmental regulations
• Analyze environmental impact
• Retrieve public comments if applicable
• Generate environmental assessment
• Route to supervisor for review
• If approved, generate permit; if denied, generate denial letter
• Notify applicant

This is a multi-step workflow with decision points. An agentic system can execute all of this autonomously, with human oversight at critical points.

PRACTICAL USE CASES

Use Case 1: Autonomous Case Processing

Scenario: A benefits agency receives thousands of applications. Each application must be reviewed for completeness, checked for eligibility, and a decision letter sent to the applicant.

Approach: An agentic system is given the goal "Process all new benefit applications." It:

• Retrieves all new applications
• For each application:
• Checks completeness; if incomplete, generates a request letter
• Checks eligibility using current regulations
• Generates a decision letter
• Routes all generated correspondence to supervisors for review before sending
• Logs all actions for audit purposes

Benefit: The system processes routine applications autonomously, freeing staff to focus on complex cases.

Risk: The system might misinterpret applications or regulations. All decisions must be reviewed before affecting the applicant.

Use Case 2: Automated Research and Analysis

Scenario: A policy office needs to research how other states handle a particular issue.

Approach: An agentic system is given the goal "Research how other states handle [issue]." It:

• Queries a database of state regulations and policies
• Retrieves relevant documents from each state
• Extracts key information
• Synthesizes an analysis
• Generates a report

Benefit: The system completes research faster than a human would.

Risk: The system might miss relevant documents or misinterpret them. The analysis must be reviewed by subject matter experts.

ANTI-PATTERNS

Anti-Pattern 1: Over-Automating Without Human Oversight

RISK: The system makes decisions autonomously without human review, so errors are made at scale.

HOW TO AVOID

-- For agentic systems, always have human-in-the-loop at critical decision points

-- Don't automate decisions that affect citizens without human review

-- Start with low-stakes automation (routine administrative tasks) before high-stakes (benefit decisions)

Anti-Pattern 2: Trusting Tool Outputs Without Validation

RISK: A tool returns unexpected information. The system uses it anyway, leading to wrong decisions.

HOW TO AVOID

-- Design tools to return only expected types of information

-- Have the system validate tool outputs before using them

-- Monitor system behavior to detect when tools are returning unexpected information

Anti-Pattern 3: Loss of Accountability Through Autonomy

RISK: Because the system acts autonomously, no one is responsible for decisions.

HOW TO AVOID

-- Maintain clear accountability: someone is responsible for the agentic system's decisions

-- Log all actions the system takes for audit purposes

-- Have humans review significant decisions the system makes

KEY TAKEAWAYS

• Agentic AI systems act autonomously toward goals, breaking problems into steps and executing them.
• Reasoning and chain-of-thought make AI systems' thinking visible, enabling verification and debugging.
• Tool use lets AI systems access external systems (databases, regulations, calculations), grounding their decisions in current information.
• Multi-step workflows enable agentic systems to handle complex processes autonomously.
• Agentic systems are powerful but require careful governance: human oversight, tool validation, accountability, and audit trails.
• Errors from agentic systems can affect many people at scale, so human review is essential before autonomous deployment.

TERMS AND GLOSSARY

• Agentic AI: An AI system that acts autonomously toward goals, planning and executing multi-step workflows without being explicitly told each step.
• Chain-of-Thought: A reasoning technique where the AI system shows step-by-step reasoning, making its thinking visible.
• Tool Use: Capability for an AI system to call external tools (functions, databases, APIs) to access information or take actions.
• Autonomous Agent: An AI system given a goal and allowed to plan and execute steps autonomously to achieve that goal.

You've completed Lecture 2.1.8, the final lecture of Chapter 1. You now understand the full range of modern AI technologies: from basic supervised/unsupervised/reinforcement learning paradigms, through transformers and LLMs, through generative AI and multimodal systems, to emerging agentic capabilities.

Chapter 2 shifts focus. You've learned the "what" and "how" of AI. Now we focus on the "where" and "why" -- finding opportunities to use AI in your agency's workflows and using AI tools effectively.

Government AI CLUB Certification Program

Level 2: AI Ready | Emerging AI Capabilities: Agents, Reasoning, and Tools | Lecture 2.1.8

A GOVT.CLUB initiative.

<- 2.1.7 Multimodal AI: Text, Image, Audio, Video
2.2.1 Workflow Analysis: Finding AI Opportunities ->

Start Your CLUB Certification

This lecture is part of L2: AI Practitioner -- 40 hours of comprehensive government AI training.

Explore CLUB Certification

Related Lectures

L2
2.1.1 -- Supervised vs. Unsupervised vs. Reinforcement Learning
60 min - Video + Interactive

L2
2.1.2 -- How Transformers and LLMs Work
60 min - Video + Diagrams

L2
2.1.3 -- Generative AI Deep Dive
60 min - Video + Demos