TOPIC 9.9

The Rise of Agentic AI

⏱️35–45 min read
📚Research

Topic 8.9

The Rise of Agentic AI

Agentic AI marks a shift from language models that respond to systems that act: planning steps, calling tools/APIs, coordinating workflows, and persisting state to achieve outcomes. For the digital economy, this is not a “feature upgrade”—it changes the unit of automation, the boundary of the firm, and the governance surface area.

Learning Objectives

By the end of this chapter, students should be able to:

  1. Explain what makes an AI system “agentic” (vs. a chatbot).
  2. Identify the technical enablers (tools, memory, orchestration, evaluation).
  3. Analyze economic implications: productivity, labor, market structure, and new intermediaries.
  4. Diagnose the main risk categories (security, manipulation, reliability, liability).
  5. Propose governance and design choices for safe deployment in public and private sectors.

1) From “Chat” to “Do”: What Makes an Agent

A useful working definition for this course:

An agentic AI system is a model (or model ensemble) embedded in a loop that can (a) set or refine goals, (b) plan steps, (c) call tools, (d) observe results, (e) update state/memory, and (f) continue until completion or handoff.

In practice, agentic systems look like:

  • Tool-using assistants: query data, draft outputs, run calculations, schedule tasks.
  • Workflow orchestrators: coordinate multi-step business processes across systems.
  • Multi-agent teams: specialized agents (research, compliance, execution) coordinating.

Core Components (Conceptual Stack)

  1. Model: the reasoning-and-language engine.
  2. Tools: APIs for actions (search, CRM, payments, procurement, code execution).
  3. State / Memory: short-term context and longer-term records (preferences, history, policies).
  4. Orchestration: a controller that manages turns, tool calls, retries, and stopping rules.
  5. Environment: data sources + real systems (internal apps, web services, humans).
  6. Evaluation & Guardrails: tests, policies, and runtime checks to prevent failure modes.

2) Why Agentic AI Is Rising Now

Agentic AI becomes viable when three conditions converge:

(a) Tool Interfaces Are Becoming Standard

Modern software is increasingly API-first, enabling agents to:

  • Read/write to CRMs, ERPs, ticketing systems, content platforms
  • Trigger workflows (webhooks, queues)
  • Execute structured actions with permission boundaries

(b) Models Became Reliable Enough for “Semi-Structured” Work

Even when models remain imperfect, they are often good enough to:

  • Propose plans that can be verified
  • Fill forms and transform data
  • Draft outputs for review

The key economic change is not perfect autonomy; it’s lower coordination cost for complex workflows.

(c) Orchestration Patterns Matured

The rise of practical patterns—tool-calling, structured outputs, retrieval, and evaluation—reduces the failure rate of “do things in the world” systems.

A masters-level takeaway:

Agentic AI is best understood as a systems engineering and governance problem, not only a model capability problem.

3) Digital Economy Impacts: Where the Value Shifts

3.1 Productivity and the “Workflow Unit”

Classic automation targets tasks (a single step). Agentic AI targets workflows (a sequence of steps with branching, verification, and handoffs).

Implications:

  • Higher ROI from automating multi-step processes (procurement, onboarding, claims)
  • New bottlenecks emerge: permissions, data quality, human review capacity
  • Competitive advantage moves toward process design and tool ecosystems

3.2 Labor Market: From Tasks to Roles and Supervision

Agentic systems change roles in two directions:

  • Augmentation: professionals manage higher-volume pipelines (more cases per analyst)
  • Restructuring: new roles (agent supervisor, workflow designer, AI auditor)

A useful frame for students:

  • The “job” becomes a bundle of workflows.
  • Workflows can be partially automated.
  • The residual human work shifts toward review, exception handling, and governance.

3.3 Market Structure: New Intermediaries and Platform Power

As agents transact and coordinate:

  • APIs become marketplaces (distribution shifts from “apps” to “agent toolchains”).
  • Standards become power (who defines schemas, permissions, identity, payments).
  • Data access becomes leverage (agents amplify the value of clean, well-governed data).

3.4 Public Sector and Development

For national digital economy strategy, agentic AI raises urgent questions:

  • How to deploy agents for public service delivery without increasing surveillance?
  • How to build local capacity (skills, data, governance) to avoid dependency?
  • How to ensure inclusion (language access, accessibility, affordability)?

4) The Risk Surface: What Can Go Wrong

Agentic systems extend risk from “bad text” to “bad actions.” Key categories:

  1. Security & Tool Abuse: prompt injection, credential misuse, lateral movement.
  2. Reliability: brittle planning, unhandled edge cases, silent failures.
  3. Manipulation & Market Abuse: coordinated agents gaming marketplaces or ad auctions.
  4. Liability: unclear accountability across model provider, tool provider, deployer.
  5. Distributional Risks: job displacement, unequal access, concentration of value.

A practical evaluation rule:

If an agent can call a tool, then the tool becomes part of the threat model.

5) Governance and Design Principles (What “Good” Looks Like)

5.1 Principle: Constrain Capabilities by Default

  • Least-privilege permissions
  • Allow-lists for tools and domains
  • Explicit approval for irreversible actions (payments, deletions, compliance filings)

5.2 Principle: Make Behavior Inspectable

  • Durable logs of: prompts, tool calls, results, and decisions
  • Traceability for audits and incident response
  • Clear handoff points to humans

5.3 Principle: Build Evaluation Into the Lifecycle

  • Pre-deployment tests (golden tasks)
  • Runtime monitoring (failure rates, anomaly detection)
  • Post-incident learning (root-cause analysis + guardrail updates)

6) Case Vignettes (Digital Economy Context)

Choose one vignette and analyze it with the frameworks above:

  1. SME Export Assistant: finds buyers, drafts outreach, checks compliance, generates invoices.
  2. Public Benefits Intake: guides applicants, checks eligibility, requests documents, schedules follow-ups.
  3. Local Procurement Agent: compares vendors, negotiates terms, drafts contracts, flags risk.
  4. Media Integrity Monitor: monitors misinformation, generates alerts, supports response teams.

For each vignette, specify:

  • Tools needed (APIs)
  • Data dependencies
  • Failure modes
  • Governance (human approvals, audit requirements)

7) Embedded Video (Local)

This platform supports embedding locally hosted MP4s from public/videos/.

Video: From Atoms to AI (Context for the Agentic Shift)

Watch prompt: Identify where the narrative shifts from infrastructure to systems. What additional layers are required for agency (tools, permissions, memory, governance)?

8) Instructor Playlist (Add 2–4 External Videos)

To embed external videos (YouTube/Vimeo) in this chapter, paste an iframe here. Keep each clip to ~10–20 minutes and pair it with a short prompt.

Template:

<iframe
  width="100%"
  height="420"
  src="PASTE_EMBED_URL_HERE"
  title="Video title"
  frameborder="0"
  allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share"
  referrerpolicy="strict-origin-when-cross-origin"
  allowfullscreen
></iframe>

Recommended clip themes (search cues):

  • “LLM tool use / function calling” (how agents call APIs)
  • “Prompt injection / agent security” (new threat models)
  • “Multi-agent workflows in enterprise” (orchestration patterns)
  • “AI governance and accountability” (liability and audit)

Discussion Questions

  1. Which sectors in your country/region have the highest workflow-automation potential—and why?
  2. Does agentic AI increase or decrease market concentration in your domain? Explain.
  3. Where should liability sit when an agent makes a harmful decision (model provider, deployer, tool vendor, user)?
  4. What are the minimum governance requirements for deploying agents in public services?

Mini-Assignment (30–45 minutes)

Pick one workflow from your university or workplace (e.g., travel reimbursement, admissions triage, vendor onboarding). Produce a one-page “agent spec”:

  • Goal and success metrics
  • Required tools and permissions
  • Data needed (and what must be redacted)
  • Human review points
  • Logging and audit plan
  • Failure-handling strategy