Agents

What's Happening in AI?

Student presentation — Week 6

Administration

• AI Gallery
• High-level feedback on Portfolio Drafts
• Review of Class Goals
• Lab today?

Course Evaluations

• 10 minutes for course evaluations

Agents

An agent is anything that can be viewed as perceiving its environment through sensors and acting upon that environment through actuators.

• Russel & Norvig 1995

Agents: Core Concepts (Huyen, 2025)

• An agent is characterized by:
  • The environment it operates in
  • The set of actions it can perform
• AI is the "brain" that:
  • Processes tasks
  • Plans sequences of actions
  • Determines task completion

The terminology of 'agents' has been muddled in AI community - there are many things that people call agents

We'll focus on large language models that can plan the necessarily actions to complete a task, have access to one or more tools to execute those actions, and summarize or use the outputs of those actions

Exercise: Planning a Task

Agent Components

• Environment: Defines where the agent operates (web, code editor, game, etc.)
• Tools: Extend the agent's capabilities
  • Knowledge augmentation - add access to source documents, rather than relying on model's learned nowledge (e.g. retrieval augmented execution)
  • Capability extension - the things that Language Models can't do themselves (e.g. calculator, code execution, image generation)
  • Write actions (file editing, sending emails)
• Planning: The agent's ability to reason about how to accomplish tasks

Huyen, 2025

We've Seen Agents in Class Already

• Retrieval-Augmented Generation -> Giving an LLM access to web search, or search over your own database
• DALL-E -> Currently, GPT doesn't generate images itself - If you ask for an image, it generates a text-to-image prompt and sends that prompt to a separate image generation tool
• Code Interpreter -> In our class about classification, we saw that you can visual your data at the end of a (e.g. our 'sentiment classification' exercise). There, the LLM is writing the code to visualize the data, then sending that code to a separate tool to run, then returning and interpreting the result.

The Future of Agents

• Autonomous assistants for complex tasks
• Potential to save human time despite computational costs
• Continuous improvement through:
  • Better planning capabilities
  • More sophisticated tools
  • Enhanced security measures

The risks of agents

• Real-world harm
• Unauthorized and highly-sensitive actions (financial transactions, data deletion)
• Privacy violations through excessive data access
• Security vulnerabilities from tool misuse
• Resource depletion (lots of read/write)
• Accountability gaps (who's responsible when agents make mistakes?)
• Amplification of existing biases in decision-making
• Unpredictable behaviors in complex environments

Agents Need More Powerful Models

• Compound mistakes: Errors multiply across multiple steps, so agents need
  • 95% accuracy per step → 60% accuracy over 10 steps
• Higher stakes: With tools, agents can perform more impactful tasks
  • You likely don't want to give access to your email or bank to any machine, but especially not to a poor, error-prone model

Agent Failure Modes

• Planning failures: Incorrect reasoning, missing steps
• Tool failures: Incorrect tool selection, misuse of tools
• Efficiency issues: Too many steps, redundant actions

Limits of red teaming

Since the model has access to the internet, the external red teamers were advised to avoid prompting the model to complete tasks that could cause real-world harm. In certain cases, they created test environments—such as mock websites, databases, or emails—to safely demonstrate possible exploits. Given this constraint, their findings may not fully capture the worst-case real-world risks, but still identified key vulnerabilities that informed additional mitigations

• OpenAI - Operator System Card

Agents-lite

Code-using LLMs are a basic - but powerful - form of tool-use agent

• see 'code interpreter' in ChatGPT, Javascript artifacts in Claude

Langchain

Problem: Limits to bespoke tools

Some tools are very valuable, like web search or being able to run Python code. Other tools are easy to implement, like a calculator.

But there's a whole lot of actions you may want to do, and building a tool for each one can get challenging.

Solutions?

Computer Use

• Claude Computer Use
• OpenAI Operator
• Google Mariner

Deep Research

Claude Code

Aider

Browser Use

Coding Agents: AI That Writes, Runs, and Ships Code

From Autocomplete to Autonomy

The history of AI coding assistance has three phases:

What Can Coding Agents Actually Do Now?

Modern agentic coding tools can:

• Read your entire codebase — not just one open file
• Edit multiple files to implement a feature end-to-end
• Run tests and fix failing ones iteratively
• Execute terminal commands (install packages, run linters, build)
• Commit code and push to GitHub
• Spawn sub-agents to work on different parts of a task in parallel
• Integrate with CI/CD — review pull requests, catch regressions

This is not science fiction — it is what Claude Code, Cursor, and Aider do in production today.

Kolterhoff, RedMonk, 2025; ikangai.com coding tools guide, 2025

The Coding Agent Landscape (2026)

Sources: Sacra/TechBuzz (Cursor ARR Feb 2026); Anthropic (Claude Code adoption); GitHub/Microsoft (Copilot stats, July 2025)

"Vibe Coding"

A new term for a new kind of development

What Is Vibe Coding?

"There's a new kind of coding I call 'vibe coding', where you fully give in to the vibes, embrace exponentials, and forget that the code even exists."

— Andrej Karpathy, February 2025

Coined by Karpathy (co-founder of OpenAI, former AI lead at Tesla), the term describes:

• Describing what you want in plain language
• Accepting AI-generated code without reviewing it line by line
• Relying on results and follow-up prompts to guide changes
• Iterating conversationally until the software does what you want

Collins English Dictionary Word of the Year 2025

Karpathy, X (formerly Twitter), February 2025; Collins Dictionary, 2025; Wikipedia: Vibe Coding

Vibe Coding: Opportunity and Caution

For LIS students: vibe coding is most valuable for small, high-value, low-risk tasks — not for systems that handle sensitive patron data.

Claude Code: A Deep Dive

Claude Code (launched February 2025 as research preview; general availability May 2025):

• Terminal-based: runs in your project directory, reads everything
• Whole-codebase context: not just the file you have open
• Agentic loop: proposes changes → executes with approval → checks output → iterates
• CI/CD integration: can commit, push to GitHub, review pull requests
• VS Code extension: added September 2025; also web and iOS versions

Internally at Anthropic: 20% adoption on day one, 80%+ of engineers using it daily within a week of internal release.

Anthropic blog; Claude Code origin story, Panaversity/Agent Factory; Medium: Evolution of Claude Code, 2025

Claude Code in Action

Reads the codebase, proposes and implements changes, runs tests — all from the terminal

Why Did It Get Significantly Better in Late 2025?

Claude Code's quality is tied to the underlying model:

• February 2025 — launched with Claude 3.7 Sonnet (first hybrid reasoning model; set state-of-the-art on SWE-bench for real software issues)
• May 2025 — Claude 4 (Sonnet 4 + Opus 4) released alongside Claude Code GA; Cursor called Opus 4 "a leap forward in complex codebase understanding"
• Late 2025 — Claude 4 model family matures; parallel multi-agent execution added; $1B annualized revenue reached by November 2025

The key insight: Compound reasoning matters more in agentic tasks than in single-turn chat. Each step in a multi-file refactor must be correct. Better models dramatically change what's possible.

Anthropic: Claude 3.7 Sonnet and Claude Code (Feb 2025); Introducing Claude 4 (May 2025)

OpenAI Codex: Cloud-Based Autonomous Coding

Codex (OpenAI, May 2025 preview / June 2025 GA) — a coding agent built into ChatGPT:

• Cloud-based: runs in OpenAI-managed sandboxes — no local setup required
• Asynchronous: assign a task, come back when it's done — like delegating to a junior engineer
• Parallel: multiple tasks can run simultaneously in separate sandboxes
• GitHub-native: reads your repo, writes code, proposes pull requests

The key contrast with Claude Code: Claude Code reasons with you in real-time in your terminal; Codex works silently in the background and surfaces results.

OpenAI: Introducing Codex, May 2025

Claude Cowork: Agents for Everyone

Claude Cowork (Anthropic, January 2026) — agentic AI without a terminal:

• Desktop agent: built into the Claude macOS app — no coding knowledge required
• Folder sandbox: designate a folder; Claude reads, creates, and edits files there
• Plain-language tasks: describe what you want; Claude plans and executes autonomously
• Computer use (March 2026): can open apps, navigate browsers, fill in forms directly on your Mac

Example tasks: "Organize my downloads folder by type and date", "Make a spreadsheet from these receipts", "Draft a report from these scattered notes"

Claude Code is for developers. Cowork brings the same agentic loop to information professionals — including LIS students.

Anthropic: Introducing Cowork, January 2026

Coding Agents for Information Professionals

You don't have to be a software engineer to benefit

LIS-relevant use cases where coding agents reduce friction:

The Democratization Argument — and Its Limits

The case for optimism:

• 81% of developers now use AI coding assistants (2025)
• Coding agents may extend coding-adjacent capability to people who couldn't justify learning full programming
• Particularly relevant for one-off analysis tasks (not maintaining production systems)

The honest counterweight:

• "Vibe coding" for critical library systems, patron databases, or systems with security requirements is risky
• Prompting effectively takes practice — there is a learning curve
• Debugging AI-generated code you don't understand can be harder than starting over
• The "10x productivity" claims assume you know enough to evaluate the output

The Week 5 framing still holds: think of structured formats (SQL, Python, R) you already use. Coding agents lower the barrier to those formats — they don't eliminate the need to understand what you're asking for.

Connection: Week 5 → Week 6

In Week 5, we saw coding assistance as a form of structured-language help for professionals who aren't programmers.

What's changed since then:

The shift is real — but it amplifies both capability and risk.

Lab Concept: Try a Coding Agent

Goal: Experience agentic coding firsthand — even without prior programming experience

Options (choose one):

Exercise: Ask an Agent About This Course

A live demo of agents as document-understanding tools — using AI to explore the course itself.

This is RAG + agents in action: the agent reads the corpus to answer questions, rather than guessing from training memory. This is exactly how enterprise document search and library knowledge systems are being built.

Code Interpreter / Claude Artifacts

The Big Quiz

See Canvas. We'll do the quiz in pairs. Remember - it's just review, unless you include in your Portfolio.

Updates

Updates since I last taught the class are here. They are not formatted as slides, so refer to the 'Notes' version of the deck https://ai.porg.dev

MCP

Model Context Protocol is a standard for providing context to tools to LLMs.

https://modelcontextprotocol.io/

e.g. You can make a database or a filesystem available to a coding app, or a chat app, etc., without writing a custom tool for each

(Mar 2025 - OpenAI is introducing MCP support: https://openai.github.io/openai-agents-python/mcp)

Playwright-MCP

• Playwright is a browser-testing tool, it's what Browser Use uses
• with this server, you can give a command to an MCP LLM (e.g. Claude on your computer) and it can use a browser (in image-based-mode or all text-based)

(Simon Willison summarizes the tool and how to use it in Claude: https://simonwillison.net/2025/Mar/25/playwright-mcp/)