AIRQ Framework ยท Agent Classification

Agent Classification

The ten-class taxonomy that organises the AIRQ landscape — scope, design principles, and classification rules for placing any agentic product.

1 Purpose

AIRQ classifies agent products along a primary taxonomy so that risk findings can be written at the class level rather than per-vendor. The taxonomy is optimized for:

  • Unambiguous classification — a new agent can be placed without debate.
  • Universal readability — terms work for technical, security, VC, and executive audiences.
  • Attack-surface coherence — members of a class share a broadly similar AIRQ-01–AIRQ-10 profile so class-level findings are meaningful.
  • Scope alignment — only end-user agent products are classified; agent infrastructure (frameworks, retrieval APIs, browser-as-a-service, vector DBs) is out of scope.

2 Design Principle

Every class is defined by its Scope — a one-line statement of what the agent does and where it acts. Sub-groups split a class along a single dimension (autonomy level, hosting model, delivery model, interaction paradigm, corpus type, or openness) where that dimension meaningfully changes product behavior or risk profile.

3 Taxonomy Overview

The ten classes, their scopes, and the number of sample agents per class.

#ClassScopeSamples
1General Chatbot AgentsChat with a general-purpose AI assistant20
2Work Copilot AgentsHelp one worker inside their tools5
3Coding AgentsWrite and modify code26
4Browser AgentsDrive the browser10
5Computer AgentsDrive the whole computer13
6Deep Research AgentsFind and synthesize the answer15
7Internal Search AgentsRetrieve and navigate internal information10
8Real-Time Voice AgentsOperate through calls, speech, and real-time audio5
9Custom Workflow AgentsCompose tool calls into user-built flows — workflow automation15
10Business Process AgentsRun an assumed business process end-to-end — process orchestration49
Total168

4 Class Definitions

Each class defined by a scope sentence followed by its sub-groups and sample agents.

4.1 General Chatbot Agents

Scope: Chat with a general-purpose AI assistant — standalone conversational destinations that increasingly use tools, memory, and MCPs. Distinct from Work Copilot Agents in that the chatbot is the product itself, not an assistant embedded in a work surface.

  • General-purpose SaaS chatbots — ChatGPT, Claude, Google Gemini, Microsoft Copilot (consumer), Grok, Meta AI, Manus AI, Poe, DeepSeek, Mistral Le Chat, Qwen Chat
  • Self-hosted chatbot frontends — LibreChat, Open WebUI, Jan, AnythingLLM, LM Studio, Msty, Chatbox AI, BionicGPT, GPT4All

4.2 Work Copilot Agents

Scope: Help one worker inside their tools — assist a single user embedded inside their daily work surface (email, documents, chat, meetings, CRM view). Acts for the individual, over that individual's data and permissions.

  • Embedded copilots — Microsoft 365 Copilot, Google Gemini Workspace, Slack AI, Notion AI, Zoom AI Companion

4.3 Coding Agents

Scope: Write and modify code — operate on source code, repos, IDEs, and CLIs.

  • Coding copilots (IDE-embedded, user present per suggestion) — Cursor, GitHub Copilot, Sourcegraph Cody, Continue.dev, Tabnine, Windsurf, Cline, RooCode, JetBrains Junie, Augment Code, Google Antigravity, PearAI
  • Autonomous coding agents (goal → repo) — Devin, OpenAI Codex, Claude Code, Aider, SWE-Agent, Amazon Q Developer, Jules, Amp, OpenHands, Goose
  • App builders (spec → deployable app) — Replit Agent, Lovable, v0 by Vercel, Bolt

4.4 Browser Agents

Scope: Drive the browser — operate web pages via navigation, forms, clicks, extraction, and page-grounded actions. Scope is one or more browser tabs.

  • General-purpose browser control — OpenAI Operator, Google Project Mariner, Skyvern, Browser Use, NanoBrowser
  • Chat-with-browser (agentic browsers with built-in chat) — Perplexity Comet, ChatGPT Atlas, Dia, Brave Leo, Opera Aria

4.5 Computer Agents

Scope: Drive the whole computer — operate local machine, desktop apps, filesystem, and communication surfaces. Distinguished from Browser Agents by full OS access.

  • Vendor-managed computer-use — Claude Computer Use, Claude Cowork, ChatGPT Agent, Perplexity Computer
  • Self-hosted computer agents — AutoGPT, OpenClaw, NemoClaw, Leon, Hermes (Nous Research), Open Interpreter, Agent S, OpenAdapt, NanoBot

4.6 Deep Research Agents

Scope: Find and synthesize the answer — ingest sources (web, papers, user-uploaded docs, structured data) and return a cited synthesis. The output is a report or dataset, not an action taken in a user system.

  • Open-web research — Perplexity, OpenAI Deep Research, Gemini Deep Research, You.com Research Agent, Genspark, GPT Researcher, Open Deep Research, Local Deep Research
  • Analysis & scientific research — NotebookLM, Elicit, Consensus, Scite Assistant, OpenEvidence, SciSpace, ResearchRabbit

4.8 Real-Time Voice Agents

Scope: Operate through calls, speech, and real-time audio — the primary interaction channel is voice. Primary channel wins even when the underlying behaviour is support, sales, or workflow.

  • Managed voice agents (vendor ships the complete agent for a specific role) — PolyAI, Parloa
  • Voice agent builders (customer builds the voice agent using the platform’s SDK / studio) — Retell, Vapi, Bland

4.9 Custom Workflow Agents

Scope: Compose tool calls into user-built flows — no-code / low-code workflow automation where the customer composes a custom flow across arbitrary apps. Execution is event-triggered; state is ephemeral; the flow is authored by the customer, not shipped by the vendor.

  • Closed-source — Zapier AI Agents, Make.com, Workato, Lindy AI, Gumloop, Relay.app, Tray.ai, Integrately, Retool Agents
  • Open-source — n8n, Activepieces, Pipedream, Dify, Langflow, Flowise

4.10 Business Process Agents

Scope: Run an assumed business process end-to-end — vendor-shipped agents whose job is process orchestration across a multi-step domain workflow with persistent state and human handoffs. The process is given (support, sales, security investigation, ITSM, SRE); the agent fills in the details. This class spans verticals — it is a horizontal substrate class that happens to productise many vertical operations, not a vertical class in disguise.

  • Process-aligned agents (vendor ships the role, customer tunes it)
    • Customer support — Sierra, Ada AI, Intercom Fin, Zendesk AI Agents, Decagon, Freshworks Freddy AI Agent, DocsBot, Gorgias, Kustomer, Tidio Lyro, Front, Yuma AI, Chatwoot
    • Sales & CRM — 11x, Artisan, Claygent, Salesforce Agentforce, HubSpot Breeze
    • Enterprise suites — SAP Joule, Workday Illuminate
    • Security & SRE — Dropzone, Prophet Security, 7AI, Simbian, Cleric, Resolve AI, PagerDuty AI, Datadog Bits, Console
    • ITSM & employee support — Moveworks, ServiceNow AI Agents, Atomicwork, Aisera, Freshservice Freddy AI, Atlassian Rovo Service, BMC HelixGPT, SysAid AI Agents, ManageEngine ServiceDesk Plus AI, Rezolve.ai, Siit, eesel AI, monday service
  • Agentic process builders (substrate on which customers build their own business-process agents) — Amazon Bedrock Agents, UiPath AI Agents, Microsoft Copilot Studio, Gemini Enterprise, IBM watsonx Orchestrate, Kore.ai Agent Platform, Oracle AI Agent Studio

5 Classification Rules

5.1 General Rules

One product, one class

A single product is classified into exactly one class, regardless of mode, template, or capability breadth. Classify by what the vendor canonically ships. A vendor may have multiple distinct products that land in different classes — that is normal (e.g. OpenAI ships ChatGPT, Deep Research, Operator, ChatGPT Agent, and Codex across five classes).

Cross-cutting dimensions

These dimensions apply on top of any class and can be used for secondary filtering, but do not create new classes:

  • Channel — chat, voice, email, API
  • Mode — copilot, autonomous, background, scheduled
  • Audience — consumer, developer, enterprise-internal, enterprise-external
  • Openness — closed SaaS, open-source, self-hosted, IDE plugin
  • Vertical — legal, medical, sales, recruiting, finance, etc.

5.2 Class Boundaries

Several class boundaries are near-neighbors and worth explicit guidance to prevent misclassification.

General Chatbot vs Work Copilot

  • General Chatbot Agents are standalone destinations (the user goes to chatgpt.com).
  • Work Copilot Agents are embedded in a work surface (the copilot lives inside Outlook, Slack, Notion, etc.).

Deep Research vs General Chatbot

  • General Chatbot Agents are the historical baseline — the general-purpose conversational destination. Many specialised classes (Deep Research, Browser, Computer, Coding) originated as features inside general chatbots.
  • Deep Research Agents produce a distinct output (cited synthesis report, not conversational reply) with a distinct attack-surface profile (multi-source ingestion, citation integrity, longer autonomous execution). When a chatbot feature becomes a distinct use case with its own output type and risk profile — as vendors and analysts now treat deep research — it warrants separate classification.

Coding Agents vs Work Copilot

  • Coding Agents operate exclusively within the code ecosystem (IDE, repo, CLI, CI/CD). The substrate is source code and developer tooling; permissions are repository-scoped.
  • Work Copilot Agents operate across general-purpose business tools (email, documents, chat, calendar, CRM). The substrate is the employee’s daily work surface; permissions are identity-scoped.

Both are embedded in a work surface and assist a single user — the distinguishing axis is substrate specificity, not embedding.

Work Copilot vs Business Process

  • Work Copilot Agents act for one employee inside their tools (M365 Copilot assists Alice with her email).
  • Business Process Agents act for the organisation on an assumed process (Agentforce qualifies all inbound leads).

Custom Workflow vs Business Process

  • Custom Workflow AgentsWorkflow automation: the flow builder lets the customer compose a custom process. The vendor ships the builder; the customer authors the flow; state lives only for the execution.
  • Business Process AgentsProcess orchestration over an assumed process (sales outreach, support, security investigation, ITSM). The vendor ships the operator; state persists across turns; human handoffs are native.

Internal Search vs Deep Research

  • Internal Search Agents retrieve and navigate inside a closed corpus; the output is documents and passages.
  • Deep Research Agents synthesise a cited answer across open or user-supplied sources; the output is a structured report or dataset.

Browser vs Computer

  • Browser Agents are scoped to a browser tab.
  • Computer Agents have full OS access (filesystem, local apps, network).

Voice vs General Chatbot vs Business Process

  • Real-Time Voice Agents — primary interaction channel is calls, speech, or real-time audio. Primary channel wins even if the underlying behaviour is support, sales, or workflow.

5.3 Out of Scope

AIRQ classifies end-user agent products. The following are agent components or code-only libraries and are explicitly out of scope:

  • Retrieval APIs — Tavily, Exa, Serper, Kagi Search API
  • Enterprise search infrastructure — AWS Kendra, Elastic AI Search Applications, Algolia NeuralSearch
  • Browser infrastructure — Browserbase, Steel, Steel Browser, Playwright MCP, Stagehand
  • Scraping / data infra — Firecrawl Agent
  • Agent-builder frameworks (code-only) — LangGraph, CrewAI, AutoGen, OpenAI Assistants SDK, LlamaIndex
  • Chatbot-building frameworks (code-only) — Rasa Open Source