AI agent

Based on Wikipedia: AI agent

On November 12, 2025, The Wall Street Journal dropped a reality check that rippled through Silicon Valley: after a frenzy of deployments, fewer than 15% of companies using AI agents had seen measurable returns on investment. The report landed just weeks after Block Inc. slashed 4,000 jobs—jobs CEO Jack Dorsey publicly blamed on AI’s encroachment. Yet here’s what most missed: Dorsey wasn’t railing against chatbots or image generators. He was staring down the rise of something far more consequential—AI agents, autonomous systems quietly rewriting how work gets done. Forget the chatbots of 2022; this new breed operates like digital employees, making decisions, booking flights, debugging code, and navigating corporate software without constant human babysitting. And while Wall Street fixated on stock dips, a quiet revolution in agentic AI was already reshaping industries from Texas to Tokyo.

AI agents aren’t just smarter chatbots. They’re goal-driven systems built to tackle complex, multi-step tasks in unpredictable environments. Picture this: you tell an agent, "Plan a surprise anniversary trip for two to Kyoto next month under $5,000, including flights, a ryokan stay, and a Michelin-starred dinner." A chatbot might dump a list of links. An AI agent? It’ll scour real-time flight APIs, cross-reference hotel reviews, negotiate with booking engines, and email you a polished itinerary—all while adapting to sudden typos or budget changes. Its magic lies in autonomy: once given a goal, it acts without hovering over your shoulder. This isn’t sci-fi. By mid-2025, systems like OpenAI Operator (which browses the web like a human) and Devin AI (a fully autonomous coder) were field-tested in real enterprises. The Financial Times nailed it: comparing agent autonomy to self-driving cars, most sat at Level 2 or 3—think adaptive cruise control—but niche tools like SIMA, trained to play Minecraft and No Man’s Sky, hit Level 4, handling entire game worlds solo. Level 5? Still vaporware.

The concept isn’t new. Harvard’s Milind Tambe has quipped since the 1990s that defining an "AI agent" was like nailing jelly to a wall. But in early 2024, Andrew Ng—AI’s most trusted evangelist—catapulted the term "agentic" into mainstream tech lexicon during a Stanford lecture. > "Agents shift AI from reactive tools to proactive collaborators," he declared, igniting a land rush. Suddenly, every major player was building them. Google launched Agent2Agent, Microsoft rolled out AutoGen, and ByteDance’s Coze let users design agents via chat. Frameworks like LangChain became the duct tape holding these systems together, stitching together memory, tools, and decision loops. Yet even as hype peaked, a critical flaw surfaced: agents kept hallucinating hotel bookings or crashing trading bots. Reliability wasn’t optional—it was existential.

Enter the unsung heroes: frameworks battling AI’s trust gap. AgentSpec stress-tested agents against edge cases, while GuardAgent acted as a paranoid auditor, flagging risky actions before execution. H2O.ai’s predictive models even forecasted agent failures—like smoke detectors for digital meltdowns. But the real breakthrough was memory. Early agents forgot your preferences between queries, like amnesiac interns. Tools like MemGPT and MemOS changed that, creating persistent memory banks storing every user interaction. Imagine an agent recalling your allergy to shellfish months later when booking that Kyoto dinner—no more awkward emergency room detours.

How do these systems actually think? Strip away the jargon, and it’s a ballet of reasoning and action. The ReAct pattern—short for Reason + Act—dominates: an agent thinks (“Flights to Kyoto peak in spring”), acts (checks Skyscanner), observes (“ANA has a $399 deal”), then rethinks (“But dinner reservations are scarce—adjust dates”). Add Reflexion, where agents critique their own plans using LLMs, and you’ve got systems that learn from blunders. Ken Huang, a systems architect at NVIDIA, crystallized this complexity into a seven-layer reference architecture—a blueprint now industry gospel:

The Seven Layers of Agentic AI

Layer 1: Foundation models—the brain. GPT-5 or Claude 3.5, powering reasoning. Layer 2: Data operations—the nervous system. Vector databases and RAG (Retrieval-Augmented Generation) feeding real-time data. Layer 3: Agent frameworks—the skeleton. LangChain or AutoGen wiring components together. Layer 4: Deployment infrastructure—the muscles. Cloud platforms scaling agents to millions of users. Layer 5: Evaluation—the conscience. Tools like Agentic Evaluations measuring safety and accuracy. Layer 6: Security—the immune system. Guardrails against data leaks or rogue actions. Layer 7: Ecosystem—the face. Where agents meet users via apps or web browsers.

This stack isn’t theoretical. In March 2025, the city of Kyle, Texas deployed a Salesforce AI agent for 311 services—it fielded pothole reports, rerouted garbage pickups, and slashed call-center costs by 30%. But Kyle’s win masked a brutal truth: most companies were fumbling in the dark. By June 2025, Fortune documented a pattern: enterprises experimenting with agents outnumbered those deploying them 10-to-1. The Information sliced deeper, identifying seven archetypes battling for relevance: - Business-task agents (e.g., automating SAP workflows) - Conversational agents (replacing human customer support) - Research agents like OpenAI Deep Research (which digs through PDFs for insights) - Analytics agents generating boardroom-ready reports - Coding agents like Cursor (writing 40% of new software at startups) - Domain-specific agents (e.g., legal or medical experts) - Web browser agents like OpenAI Operator (navigating sites incognito)

August 2025 brought clarity. New York Magazine crowned software development as agentic AI’s killer app. Agents like Devin weren’t just debugging—they were building entire apps from scratch. Yet by October, The Information noted a retreat: businesses narrowed focus to coding and customer support, abandoning moonshots. Why? ROI remained elusive. Agents devoured engineering hours for marginal gains. A crypto startup’s trading bot might nail meme-coin surges but crash during market volatility. Social media agents scheduling viral posts often veered into tone-deaf territory. Even Hugging Face’s free Open Deep Research agent—launched in February 2025 as an open-source rival to OpenAI’s tool—struggled with academic paywalls.

The industry’s response? Standardization. In December 2025, the Linux Foundation launched the Agentic AI Foundation (AAIF), a neutral consortium to prevent a Tower of Babel scenario. Protocols like Agent Protocol (for inter-agent chats) and Gibberlink (for tool integration) emerged as potential lingua francas. Meanwhile, multimodal agents shattered text-only limits. NVIDIA’s VIMA framework, released in late 2024, let agents analyze video feeds—spotting defects on factory lines or summarizing NFL highlights. Microsoft’s secret project? An agent trained on robotics data that manipulated software UIs and physical robots, blurring lines between digital and real.

But here’s the pivot most overlook: agents aren’t replacing jobs wholesale. They’re reshaping workflows. At a major bank, agents handle 80% of routine compliance checks, freeing analysts for complex fraud investigations. In e-commerce, they manage inventory swaps during supply chain snarls—tasks humans deemed too tedious. The Block layoffs weren’t about agents taking jobs; they were about companies deploying half-baked systems that created more work than they solved. True autonomy demands precision. When an agent books that Kyoto trip, it’s not generating content—it’s orchestrating reality. The Planner-Critic workflow exemplifies this: one agent drafts the itinerary, another ruthlessly critiques it (“Too much walking for elderly parents?”), and iterations continue until flawless. This isn’t prompt engineering. It’s collaborative intelligence.

Vendors are finally catching on. AWS’s Bedrock Agents now bake in evaluation layers, while Google’s Agent Workspace includes ROI dashboards tracking cost-per-task. The Galileo AI leaderboard—ranking agents by LLM backbone—has become essential reading for CTOs. Yet for all the progress, the human element remains irreplaceable. Agents excel at tasks with clear rules; they flounder in ethical gray zones. When Kyle, Texas’ 311 agent faced a domestic violence report, it immediately escalated to a human—no algorithmic guesswork allowed.

We stand at inflection point. The Linux Foundation’s AAIF promises transparency, but the real test is mundane: will agents finally deliver reliable value? In November 2025, Salesforce quietly reported that its government clients saw 22% faster service resolution with agents—when properly constrained. That’s the blueprint: narrow autonomy, layered safety, and ruthless focus on tasks humans hate. Agentic AI won’t replace us. But it will redefine what “work” means—one Kyoto itinerary, one bug fix, one pothole report at a time.

The revolution isn’t coming. It’s already here—and it’s politely asking for your itinerary preferences.