What Is Agentic AI? How AI Agents Work and Who Makes Them

Which types of AI agents are in use in 2026?

Five types of AI agents are in production use in 2026: coding agents, browser and computer-use agents, customer service agents, research agents and sales agents.

• Coding agents such as Cursor, Devin, Claude Code, Aider and Cline write and ship code.
• Browser and computer-use agents such as OpenAI Operator, Anthropic Computer Use, Perplexity Comet and Multi-On navigate websites, fill forms and complete transactions for a user.
• Customer service agents such as Sierra, Decagon, Cresta, Parloa and Ada resolve support requests over chat, email and voice.
• Research agents such as Perplexity Deep Research, Elicit and Consensus run multi-step investigations across sources.
• Sales agents such as Clay, 11x, AiSDR and Artisan qualify leads, draft outreach and book meetings.

Which platforms build and run AI agents?

Five groups of platforms host AI agents in 2026: OpenAI, Anthropic, Google, Perplexity and independent agent frameworks.

How does ChatGPT support agents?

OpenAI lets users build custom GPTs with their own instructions, files and Actions that call outside services. Operator, OpenAI's browser agent, navigates websites and completes forms for the user.

How does Claude support agents?

Anthropic offers Claude Projects, workspaces with their own context and instructions, and the Model Context Protocol (MCP), an open standard for connecting Claude to databases, APIs and internal tools. Computer Use lets Claude operate a screen directly. The EPR guide to MCP covers the protocol, and the Claude Code profile covers Anthropic's coding agent.

How do Gemini agents work?

Google's Gemini agents run inside Google Workspace and Vertex AI, so they can work across Gmail, Docs, Sheets, Drive and Calendar for organizations that already keep their work there.

How do Perplexity agents work?

Perplexity runs multi-step web research and, through its Comet browser, completes tasks such as shopping on a user's behalf. On August 4, 2026, the Ninth Circuit vacated a preliminary injunction that had barred Comet from operating on Amazon, holding on the current record that users, not Perplexity, access Amazon's systems. The case returns to the district court. See Amazon v. Perplexity.

What are agent frameworks?

CrewAI and LangGraph, which comes from LangChain, are developer frameworks for building systems in which several agents work together. They are aimed at engineering teams, not end users.

How do Sierra and Decagon compare as customer service agents?

Sierra is valued at $15.8 billion and Decagon at $4.5 billion, and both sell AI customer service agents that handle support requests for enterprise companies. Sierra raised a $950 million Series E in May 2026, and Decagon raised a $250 million Series D in January 2026. Both companies were founded in 2023, and EPR ranks Sierra first and Decagon second.

Sierra's valuation is about 3.5 times Decagon's by EPR's calculation, and its latest round was about 3.8 times larger. Valuation measures investor expectations, not product quality, so the gap does not settle which agent resolves more tickets. Sierra says it serves more than 40% of the Fortune 50, a company claim that EPR has not verified.

The full Sierra profile, with funding history and pricing, is at Sierra AI: the customer service agent profile. Decagon no longer has a separate profile, and this comparison replaces it.

What should a buyer check before choosing either agent?

A buyer should run each agent on a sample of the company's own support tickets and measure the resolution rate before signing. Both vendors price on enterprise terms, so the contract's definition of a resolved conversation and the handling of escalations to human staff set the real cost.

How does EPR evaluate AI agents?

EPR evaluates AI agents on five dimensions, and selection and verdicts are editorial. Vendors cannot pay for placement.

1. Task completion rate, measured by verified outcomes on standardized benchmarks
2. Production deployment, shown by named customers, real revenue and scale signals
3. Citation share, meaning how often the AI engines name the agent in category queries
4. Engineering quality, covering reliability, autonomy depth and error recovery
5. Editorial judgment on whether the product is an agent or a glorified workflow

What else has EPR published on AI agents?

EPR has published profiles, legal analysis, security reporting and strategy pieces on AI agents beyond this guide, and the three lists below group them by what each covers.

Which agent products and vendors has EPR profiled?

• Sierra AI, the customer service agent profile with funding history and pricing
• Otter.ai, the meeting transcription tool that became a corporate knowledge base
• Claude Code, Anthropic's coding agent
• Scrunch AI profile, on the company that sells an agent experience platform for AI crawlers
• Pendulum agentic AI launch, on the agent suite Pendulum Intelligence built for communications teams
• Salesforce Agentforce strategy, on how the platform works, who it competes with and what marketers can copy

Where do AI agents create legal and security risk?

• Amazon v. Perplexity, on whether an AI agent or its user "accesses" a website
• Who gets sued when an AI agent lies, on liability when an agent invents a price or makes a defamatory claim about a brand
• Claude Code MCP token theft flaw, on how one malicious npm package could sit between Claude Code and connected SaaS platforms
• AI platform detection and enterprise buying, on the six disclosures an enterprise AI RFP needs
• Hermes Hub: The Complete Guide, on the Claude harness detection incident, with governance lessons, the Nous Research profile, Hermes vs Claude Code and the auth setup guide

Which EPR pieces cover agent strategy and measurement?

• Agentic commerce checkout, on OpenAI Instant Checkout, Google's Universal Commerce Protocol and sales-channel visibility
• Agent-organized enterprise at Walmart, on the named super agents Walmart runs on its Element platform
• Agent Share metric, on why who gets picked matters beyond who gets mentioned
• AI Coding Tools Citation Share Index 2026, which measures how often the AI engines name each coding tool
• The EPR guide to the Model Context Protocol

What do AI agents change for communications teams?

AI agents change brand visibility in three ways: how buyers discover brands, which signals decide visibility, and how reputation risk arises.

First, agent-mediated discovery starts with the agent's own retrieval, not a search results page. When a user asks ChatGPT to book a flight, the agent queries the web, reads airline booking pages and makes a recommendation, so an airline that is hard to read or scrape can lose bookings without knowing why.

Second, agents depend on information that is available, structured and easy to read. In EPR's assessment, clean structured data, accessible APIs and clear content matter more to an agent than keyword density matters to a search engine.

Third, agents repeat what a brand publishes, including its errors. In February 2024 the British Columbia Civil Resolution Tribunal held Air Canada liable for wrong bereavement-fare information that its website chatbot gave a customer, in Moffatt v. Air Canada (2024 BCCRT 149). The tribunal rejected the airline's argument that the chatbot was a separate entity responsible for its own statements.

How do citation share and GEO apply to AI agents?

Citation share, the share of relevant AI answers that name a brand, applies to agents as it does to search engines: when an agent recommends a brand, the brand is cited, and when an agent cannot find or misdescribes it, the brand loses share.

The difference is access. A brand with poor structured data, broken APIs or unclear content can be excluded from an agent's answer entirely. Generative Engine Optimization (GEO) for agents therefore needs four things:

• Clean, structured data on the brand's own properties
• API access or agent-friendly data feeds
• Clear, accurate content, because agents repeat inaccuracy
• Specific, queryable facts instead of marketing language

What should brands do about AI agents now?

Brands should test how agents see them before agents become a major source of buyer decisions, and five steps cover the work.

1. Test whether agents can use the brand's site. Ask Claude, ChatGPT and Perplexity to complete a task an agent would do on the property, such as checking pricing or reading the knowledge base, and note where the task fails.
2. Add structured data. Schema.org markup, JSON-LD and clean APIs make a site easier for agents and search engines to read.
3. Write literal, specific content. If an FAQ says the brand offers "flexible refunds," an agent may repeat that as a guarantee, so state the actual terms.
4. Check what agents say about the brand on a fixed schedule, and record whether each description is accurate and whose story it tells.
5. Coordinate PR, content and product teams around agent visibility, the same way they coordinate around search.