What are AI agents and how do they work?

An AI agent is software that perceives context, plans actions, calls tools (search, code, APIs, files), and pursues multi-step goals with minimal hand-holding. Unlike a chatbot — which waits for a prompt and produces text — an agent acts with intent. It decomposes the request into steps, dispatches each step to the best model and tool, writes intermediate results to shared memory, and only returns when the goal is met or a human-in-the-loop gate is hit. By 2026 most agents are built on the same three primitives: a planner LLM, a tool-call protocol (often MCP), and an evaluator to catch failures before they reach the user.

Which companies are building AI agents?

Every major lab and platform: Anthropic shipped Claude Computer Use in October 2024 and the open Agent Skills standard in December 2025. OpenAI launched Operator in January 2025. Google announced Gemini 3.5 Flash and Gemini Spark at I/O 2026 in May, positioning agents as the successor to chatbots. Meta is racing a personalized everyday assistant. Zendesk launched an Autonomous Service Workforce priced on verified resolutions. Fiserv launched agentOS for banking with a projected August 2026 GA. On the platform side: LangGraph, CrewAI, AutoGen, Vellum, Salesforce Agentforce, and ZeroTwo — each with a different bet on multi-agent shape, model coverage, and price.

What's the difference between AI chatbots and agents?

Chatbots respond — agents act. A chatbot is a single turn of text-in, text-out, locked to one model. An agent is a loop: it plans, calls tools, reads the result, decides whether to keep going, and stops only when a goal condition is met. Agents have memory across steps, can route different sub-tasks to different models, and can pause for human approval before irreversible actions. Sinan Aral of MIT Sloan puts it: 'the agentic AI age is already here.' By 2026 the practical line is whether the system can complete a multi-step task without you driving it.

How much does it cost to build an AI agent?

It depends entirely on the path. (1) Building from scratch with open-source frameworks like LangGraph or AutoGen runs roughly $20,000–$80,000 in engineering plus $100–$500/mo ongoing for model APIs, vector stores, and hosting. (2) Buying a closed-vendor agent like Operator or Gemini Spark is bundled into a $20–$200/mo seat license but locks you to one model provider. (3) Using an enterprise platform like Vellum or Fiserv agentOS is custom-quoted and typically starts in the high-five to six figures annually. (4) Using a multi-model workspace like ZeroTwo Agent costs $0 to start and $29.99/mo for Pro — Pro 2x is $59.98/mo, Ultra is $120/mo. Most teams underestimate evaluation and observability, not LLM tokens.

What are the best AI agents for business?

See the ai-agents-news page for the full answer.

Can AI agents replace human workers?

Not as a wholesale swap — but they are reshaping job composition. Gartner predicts 40% of enterprise applications will feature task-specific AI agents by end of 2026 (up from <5% a year earlier) and that 60% of brands will use agentic AI for one-to-one personalization by 2028. The pattern that's emerging is replacement of repetitive sub-tasks (ticket triage, post-trade reconciliation, document review, first-pass customer responses) and amplification of higher-skill work. Zendesk's outcome-priced workforce is the clearest signal: companies are now buying 'resolved tickets' as a unit, not 'agent seats.' Roles that were 60% routine and 40% judgment are converging on 100% judgment.

Agentic AI is AI that takes initiative — it acts with intent across multiple steps, calls tools, and adapts its plan as it learns. The term distinguishes it from prompt-and-response chatbots. MIT Sloan's Sinan Aral describes it bluntly: 'The agentic AI age is already here. We have agents deployed at scale in the economy to perform all kinds of tasks.' Practically, agentic AI is the combination of (a) a reasoning LLM as the planner, (b) tool use through an open protocol like MCP (Model Context Protocol), (c) memory across steps, (d) evaluators that catch bad trajectories, and (e) human gates on sensitive actions. Gartner's 2026 Hype Cycle for Agentic AI catalogs the maturity of each piece.

How do I use AI agents for customer service?

Three options. (1) Buy an outcome-priced workforce — Zendesk's Autonomous Service Workforce launched May 2026 and charges per verified resolution across messaging, email, and voice. (2) Build a custom support agent on a multi-model workspace: a planner agent classifies the ticket, a worker agent calls your knowledge base via MCP, an evaluator verifies the answer cites a real article, and a human gate routes high-risk replies (billing disputes, regulated topics) to a human. (3) Hybrid: closed-vendor for tier-1 deflection plus a ZeroTwo-built agent for the long-tail and edge cases. Cost-wise, outcome pricing scales with resolutions, custom builds scale with model tokens — pick based on your ticket volume curve.

ZeroTwo / Agents Desk

Updated May 21, 2026

The 2026 dossier

AI Agents News: Latest Launches, Trends & the 2026 Landscape

Q: What is agentic AI?

Agentic AI is AI that takes initiative — it acts with intent across multiple steps, calls tools, and adapts its plan as it learns. The term distinguishes it from prompt-and-response chatbots. MIT Sloan's Sinan Aral describes it bluntly: 'The agentic AI age is already here. We have agents deployed at scale in the economy to perform all kinds of tasks.' Practically, agentic AI is the combination of (a) a reasoning LLM as the planner, (b) tool use through an open protocol like MCP (Model Context Protocol), (c) memory across steps, (d) evaluators that catch bad trajectories, and (e) human gates on sensitive actions. Gartner's 2026 Hype Cycle for Agentic AI catalogs the maturity of each piece.

Q: How do I use AI agents for customer service?

Three options. (1) Buy an outcome-priced workforce — Zendesk's Autonomous Service Workforce launched May 2026 and charges per verified resolution across messaging, email, and voice. (2) Build a custom support agent on a multi-model workspace: a planner agent classifies the ticket, a worker agent calls your knowledge base via MCP, an evaluator verifies the answer cites a real article, and a human gate routes high-risk replies (billing disputes, regulated topics) to a human. (3) Hybrid: closed-vendor for tier-1 deflection plus a ZeroTwo-built agent for the long-tail and edge cases. Cost-wise, outcome pricing scales with resolutions, custom builds scale with model tokens — pick based on your ticket volume curve.

A dated chronology from Claude Computer Use (Oct 2024) to Fiserv agentOS (Aug 2026 projected) — plus a market on track from $7.84B in 2025 to $182.97B by 2033 and a buildable platform to ride the wave.

TL;DR

AI agents news in 2026 is dominated by autonomous agents from Google (Gemini Spark, 3.5 Flash), OpenAI Operator, Anthropic's open Agent Skills, and Zendesk's resolution-priced workforce. The market is on track from $7.84B (2025) to $182.97B (2033). ZeroTwo Agent lets you build and supervise agents across 60+ models.

Launch ZeroTwo Agent — free See the timeline

Sources: Gartner · GrandView · Anthropic · OpenAI · Google · MIT Sloan

Section · 01

What's happening in AI agents right now (May 2026)

Two weeks of news that compressed a year of progress.

May 2026 may be the month historians point to as the moment agents stopped being a research demo and became a product category. At Google I/O on May 19, Sundar Pichai unveiled Gemini 3.5 Flash and positioned agents — not chatbots — as the next wave of Google's AI strategy. Alongside it, Google announced Gemini Spark, a general-purpose agent in the Gemini app, with information agents for Search slated to launch summer 2026.

Eleven days earlier, CNBC reported that Meta and Google had entered "agentic wars" — Meta racing a personalized everyday assistant, Google advancing a 24/7 personal agent. The same week Zendesk launched its Autonomous Service Workforce, priced on verified resolutions instead of seats — the first major SaaS vendor to make outcome pricing a category.

Beneath the consumer announcements, an open infrastructure layer kept hardening. In December 2025, Anthropic published Agent Skills as a portable open standard, with Atlassian, Figma, Stripe, Notion, and Zapier in the launch partner directory. Fiserv has confirmed agentOS for banking with a projected August 2026 general availability — already in production at Broadridge for post-trade processing.

Section · 02

The 2024–2026 AI agent launch timeline

Eight dated milestones from frontier labs and enterprise platforms — every entry links to its primary source.

8 entries · Oct 2024 → Aug 2026

Oct 22, 2024Public beta · Anthropic
Claude Computer Use ships
Anthropic releases computer-use capability inside Claude 3.5 Sonnet, letting the model navigate screens and simulate keyboard and mouse input. First production-grade general agent loop from a frontier lab.
platform.claude.com — computer-use tool
Jan 23, 2025Preview · US · OpenAI
OpenAI Operator launches
OpenAI announces Operator, an agentic AI capable of autonomous task execution via browser and desktop control. Launched as a limited preview to ChatGPT Pro subscribers in the United States.
openai.com — Introducing Operator
Dec 2025Open standard · Anthropic
Anthropic Agent Skills (open standard)
Anthropic announces Agent Skills as a portable, open standard, plus enterprise org-wide management tools and a partner skill directory featuring Atlassian, Figma, Stripe, Notion, and Zapier.
venturebeat.com coverage
May 8, 2026Industry shift · CNBC
Meta and Google enter the agentic wars
CNBC reports Meta building a personalized AI assistant for everyday tasks while Google advances a 24/7 personal agent powered by Gemini. Both companies signal agents — not chatbots — are the next consumer interface.
cnbc.com — Meta vs. Google agentic wars
May 19, 2026Google I/O 2026 · Google
Google Gemini 3.5 Flash debuts
Google unveils Gemini 3.5 Flash and positions agents — not chatbots — as the next wave. The model can independently execute coding pipelines, manage multi-step research projects, and build operating systems end-to-end.
techcrunch.com — Gemini 3.5 Flash agentic bet
May 19, 2026Google I/O 2026 · Google
Gemini Spark announced
Google launches Gemini Spark, a general-purpose AI agent inside the Gemini app that reasons across connected applications. Companion information agents for Search are scheduled to roll out in summer 2026.
blog.google — I/O 2026 announcements
May 2026Relate 2026 · Zendesk
Zendesk Autonomous Service Workforce
Zendesk launches AI agents priced on verified resolutions rather than seats, replacing deflection-based bots across messaging, email, and voice channels. Outcome-priced support is now a category.
cmswire.com — Zendesk Relate 2026
Aug 2026 (projected)General availability · Fiserv
Fiserv agentOS for banking
Fiserv launches agentOS, an operating system for agentic AI in banking. Expected general availability by August 2026 — already in production at Broadridge for post-trade, account opening, and workflow automation.
investors.fiserv.com — agentOS launch

Section · 03

Why AI agents matter now: the market and enterprise adoption

Six numbers — drawn from Gartner, GrandView Research, and independent analyst surveys — that explain why every major platform is racing to ship an agent in 2026.

$182.97B

Projected AI agents market by 2033, up from $7.84B in 2025 (49.6% CAGR).

Source: GrandView Research

40%

Of enterprise apps will feature task-specific AI agents by end of 2026 — up from <5% in 2025.

Source: Gartner

$58B

GenAI + agent-driven market shake-up against mainstream productivity tools through 2027.

Source: Gartner Strategic Predictions 2026

60%

Of brands will use agentic AI to deliver one-to-one personalized interactions by 2028.

Source: Gartner

94%

Of companies globally use AI in at least one business function; agent adoption is accelerating fastest.

Source: Azumo — AI agent statistics

39.6%

North America's share of the global AI agents market in 2025.

Source: GrandView Research

Editor's note

The two Gartner numbers that matter most are paired: 40% of enterprise applications featuring task-specific agents by end-2026 (up from less than 5% a year earlier), and a $58B market shake-up through 2027. Translation: every productivity vendor is now rebuilding its product around agentic workflows — and the budget is coming out of existing seats, not net new spend.

Section · 04

AI agents vs chatbots: what actually changed

The category shift in one sentence: chatbots respond, agents act. Everything that follows is plumbing for that one difference.

Quoted

"The agentic AI age is already here. We have agents deployed at scale in the economy to perform all kinds of tasks."

Sinan Aral · Professor, MIT Sloan · source

Capability	Chatbot	AI agent
Trigger	User prompt	Goal + context
Loop shape	Single turn	Plan → act → reflect → repeat
Tool use	Limited / none	Native via MCP, webhooks, APIs
Memory	Conversation window	Cross-step scratchpad + persistent
Model routing	One model, fixed	Best model per step
Human gate	—	Configurable for sensitive actions
Pricing logic	Per seat / token	Increasingly per outcome (Zendesk)

See our deep dive on the AI agent platform architecture for how each capability is implemented.

Section · 05

How ZeroTwo Agent fits into the news

Every entry on the timeline is locked to a single provider — Claude, OpenAI, Google, Zendesk, Fiserv. ZeroTwo Agent is the opposite bet: any of 60+ frontier and open-weight models per step, native MCP tools, multi-agent patterns, and a free tier so you can ship a prototype today.

60+ models, no lock-in
Claude Sonnet 4.6, GPT-5, Gemini 3 Pro, Grok 4, DeepSeek R1, plus open weights. Pick the model per step, not per project.
Native MCP + supervisor/worker patterns
Tools registered via Model Context Protocol — VS Code, JetBrains, Cursor, and ZeroTwo all speak it. Plan, dispatch, act, verify, deliver.
Evaluators + human-in-the-loop
Pause agents before irreversible API calls. Run outputs through model-as-judge and schema validators before they reach a user.
Cost telemetry per agent
See exactly where the spend went — per agent, per run, per step. Route cheap-and-fast steps to small models; reserve frontier for hard reasoning.

Build your first agent across 60+ models

Pricing

Start free. Scale linearly.

No quote process, no per-vendor escalation. The same pricing whether your agent calls GPT-5 once or Claude Opus 4.6 forty times.

Free

Daily agent runs across frontier models

Choosing an agent platform: the 2026 decision matrix

Four shapes the agent market sorts into. The right answer depends less on which one is "best" than on which one matches your workflow, ops bandwidth, and lock-in tolerance.

Closed-vendor agent

Operator, Gemini Spark, ChatGPT Agent

Models: Single provider only
Tool use: Vendor-curated tools
Pricing: Bundled in $20–$200/mo seat
Best for: Consumers and individual workers who want a turnkey 'do my errand' agent inside one app.

DIY framework

LangGraph, CrewAI, AutoGen

Models: Bring your own keys
Tool use: Custom — you write it
Pricing: Open source; cloud usage-based
Best for: Engineering teams who want maximum control and are happy to own the operational and evaluation burden.

Enterprise agent platform

Vellum, Fiserv agentOS, Salesforce Agentforce

Models: Multi-provider per workflow
Tool use: Native + curated connectors
Pricing: Enterprise (custom quote)
Best for: Regulated industries — finance, healthcare, support — that need evals, observability, SSO, and DPAs.

Recommended

ZeroTwo Agent

ZeroTwo

Models: 60+ frontier + open-weight
Tool use: Native + MCP-compatible
Pricing: Free; Pro $29.99/mo
Best for: Teams that want a multi-model agent workspace today — no per-vendor lock-in, no quote process.

Try ZeroTwo Agent free — Pro from $29.99

Build a multi-model agent in minutes. No credit card.

Start free

Section · 07

What to watch next: H2 2026 → 2027

The next eighteen months are where the agentic infrastructure hardens — and where the budget shifts.

Aug 2026 — Fiserv agentOS GA (projected). Production at Broadridge for post-trade, account opening, and workflow automation. First agent operating system built explicitly for regulated banking. source

End-2026 — Gartner's 40% threshold. 40% of enterprise applications featuring task-specific AI agents by end-of-year, up from less than 5% in 2025. Translation: 8x adoption inside 18 months.

Through 2027 — $58B market shake-up. Gartner's strategic prediction on GenAI + agent-driven disruption to mainstream productivity tools. Most of that budget is rebranding existing seats — not net new.

2028 — the 60% personalization milestone. Gartner predicts 60% of brands will use agentic AI to deliver one-to-one personalized interactions. Outcome pricing (Zendesk) is the leading indicator.

2033 — $182.97B market. GrandView Research projects 49.6% CAGR from $7.84B (2025) to $182.97B (2033). North America holds 39.6% share today; expect Asia-Pacific to compress that gap by 2028.

The takeaways

Five things to remember from May 2026

Agents replaced chatbots as the consumer narrative at Google I/O 2026 — Gemini 3.5 Flash and Gemini Spark are the strongest signal yet that the category has crossed.
Outcome pricing is real: Zendesk's Autonomous Service Workforce charges per verified resolution, not per seat. Expect every enterprise SaaS to follow within 18 months.
Agentic AI is now a $7.84B market growing to a projected $182.97B by 2033 (49.6% CAGR) — and Gartner expects 40% of enterprise apps to feature task-specific agents by end-2026.
The open infrastructure layer is hardening fast: MCP for tool use, Anthropic Agent Skills as a portable standard, Fiserv agentOS as the first regulated-industry operating system.
ZeroTwo Agent is the multi-model bet against single-vendor lock-in — 60+ models per step, MCP-native, free to start, $29.99/mo Pro.

Frequently asked