AI Agent Glossary: Dots, Bots, Harnesses and More Terms

Vendors describe simple ideas in complicated words, and the confusion costs buyers real money. This glossary defines thirty AI agent terms in plain English for a non-technical owner — what each thing is, why it matters to a small business, and which product the term comes from where that helps. If you are starting from zero, read our plain-language introduction for non-technical owners first; for the two most-mentioned products, see ChatGPT Dots explained and the AI agent guide.

Key takeaways

• An agent acts across your systems in multiple steps; a chatbot converses — ask vendors which one they actually sell.
• The harness, approvals, and audit trail matter more to a buyer than which model sits underneath.
• Isolation terms (virtual machine, sandbox, secure VM) all answer one question: where does the agent work, and what can it reach?
• Pricing terms (per-seat, usage pools, per-token) stack in layers — always ask for the monthly total at your volume.
• Keep this page bookmarked and check any new vendor term against it before signing.

Agent types and products

1. AI agent. Software that carries out multi-step tasks across apps and files with limited supervision — reading, deciding, acting, and reporting back. Distinct from a chatbot, which only converses.

2. Chatbot. A conversational interface that answers questions in text. Useful for FAQs and first contact, but it does not log into systems or complete work unless an agent layer sits behind it.

3. ChatGPT Dots. OpenAI's always-on agents, announced September 29, 2026 and powered by GPT-6 Astra. Each Dot gets its own cloud computer and browser, connects to thousands of apps, and is reachable via ChatGPT, mobile, Slack, and Teams, with voice calls supported.

4. Grok Bot. xAI's always-on agents, launched August 11, 2026. Bots have their own computer, work inside tools and apps, run around the clock in parallel, and can message each other and share work. Available through SuperGrok and Cursor paid plans.

5. Meta Muse. Meta's personal AI agent, launched September 2026 in the US and powered by the in-house Muse Spark model. Consumer-first: email, shopping, travel, and errands via an app, the web, and WhatsApp, free for most needs.

6. Claude Cowork. Anthropic's agentic workspace for non-developers. It uses local files, connector apps, and browser actions, runs tasks in the cloud even with the laptop closed, and is merging into the main Claude experience over time.

7. Specialist agent. An agent scoped to one business function with its own identity and limited system access — OpenAI's procurement, invoice-processing, support, and contracts Dots now in testing are the clearest example.

8. Open-source agent. An agent whose code is public and self-hostable under a license such as MIT — OpenClaw and Hermes Agent are the leading 2026 examples. The software is free; model fees, hosting, and maintenance are not.

9. Voice agent. An agent that holds spoken conversations and acts mid-call — booking, looking up accounts, routing. Gemini's 3.8 Live speech-to-speech models, priced per audio minute, are pushing this category's costs down.

10. AI receptionist. A customer-facing specialty agent that answers business calls, books appointments, and logs outcomes — narrower than a general agent, deeper on phones. It is the category Praktivo builds in.

How agents run

11. Harness. The software scaffolding around the AI model: tools, computer and browser management, approvals, logging, and result delivery. Much of what you experience as agent quality is the harness, not the model.

12. Model. The underlying AI engine (GPT-6 Astra, Muse Spark, Grok 4.7, Claude Sonnet 5.5) that reasons and writes. Vendors swap models over time; your briefs, approvals, and integrations outlast any single one.

13. Virtual machine (VM). An isolated cloud computer where the agent works instead of on your laptop. Dots run on individual cloud computers; Muse runs inside the Muse Secure VM holding both agent and personal data.

14. Sandbox. A restricted execution area that limits what the agent can reach — certain folders, no network, no sensitive systems. Ask whether file and command access is sandboxed or full before granting permissions.

15. Background task. Work the agent continues after you close the app — monitoring, research, scheduled reports. Background capability is what separates 2026 agents from chatbots; it is also what makes approvals and logs non-negotiable.

16. Connector (integration). A sanctioned link between the agent and an app — email, calendar, CRM, Drive, Slack. Dots connect to more than 4,000 apps through OpenAI's plugin ecosystem; Cowork connects via named connector apps and bundled plugins.

17. Skill (plugin). A packaged capability bundle — instructions plus connectors plus sub-agents — that teaches an agent a job. OpenClaw ships with more than a hundred preconfigured skills; Hermes agents can create new skills autonomously after complex tasks.

18. Subagent. A subordinate agent spawned for parallel work — one researches pricing while another drafts the summary. Hermes supports isolated subagents for parallel tasks; Grok Bots can message each other and share work similarly.

19. Memory. What the agent retains across sessions: preferences, facts, past decisions. Hermes builds a deepening model of the user across sessions with searchable past-session history; OpenClaw offers persistent memory. Ask what is stored, where, and how to delete it.

20. Cron (scheduler). Time-based automation: jobs that run on a schedule without being asked. Hermes has a built-in cron scheduler delivering to any connected platform; OpenClaw supports cron and background tasks the same way.

Control, safety, and oversight

21. Approvals (human-in-the-loop). Tiered permissions — automatic, notify, require approval — governing which actions run free and which pause for a human yes. Dots, Cowork, and Muse all implement variants; refuse any business agent that lacks them.

22. Least privilege. The rule that an agent gets only the access its job needs: one folder not the whole drive, read rights before write rights, its own login never a shared one. The cheapest security measure available and the most skipped.

23. Audit trail. The log of what the agent did, when, under whose approval, and touching what data. Required for incident review, staff accountability, and regulated work; Microsoft's Agent 365 builds Purview audit and eDiscovery around it.

24. Agent identity (Agent ID). A distinct login or directory identity for each agent, managed like a user account — created, permissioned, reviewed, and offboarded. Agent 365 issues Agent IDs in Microsoft Entra; smaller shops should at minimum use dedicated service accounts.

25. Read-only mode. A starting posture where the agent may look but not change: research, draft, summarize, with sends and edits held for humans. The recommended default for every pilot's first weeks.

26. Prompt injection. Malicious instructions hidden in web pages, files, or messages that try to hijack the agent's behavior. Defenses include read-only defaults, site allowlists, approval tiers, and monitoring — ask vendors what they implement.

Pricing and buying terms

27. Per-seat pricing. A flat fee per user per month — Agent 365 at $15 per user per month, Cursor Teams Standard at $40 per user per month. Predictable, but verify what usage is actually included per seat.

28. Usage pools and tokens. Consumption-based billing for the computing behind tasks: Cursor splits first-party and third-party model pools, Grok Bot usage is billed separately from Grok and Cursor plans, and API models price per million tokens in and out. Always ask for the monthly total at your volume.

29. Free tier. A $0 entry plan with capped volume and reduced controls — Muse is free for most personal needs; OpenClaw and Hermes are free software with paid running costs. Evaluate on free, pilot on the cheapest paid tier that carries real volume.

30. Governance (control plane). The oversight layer above individual agents: registries, lifecycle rules, threat protection, data policies. Agent 365 is the enterprise reference; a small business needs the same four habits at notebook scale — one owner, one access list, one approval queue, one weekly review.

Quick-compare: the terms buyers mix up

Term A	Term B	The difference
Agent	Chatbot	Acts across systems vs only converses
Harness	Model	The scaffolding and controls vs the reasoning engine
Skill	Connector	A packaged job capability vs a link to one app
Memory	Audit trail	What the agent remembers vs the log of what it did
Approval	Notification	Pauses for a yes vs informs after the fact
Free software	Free tier	$0 license with running costs vs $0 plan with caps

FAQ

What is the difference between an AI agent and a chatbot?

A chatbot produces text in response to a prompt, while an agent takes multi-step action across systems: reading files, opening browsers, updating records, and completing tasks in the background. Chatbots answer, agents do. Many products bundle both, which is why vendors blur the terms — ask whether the system can act in your apps or only converse.

What does harness mean in AI agents?

The harness is the software scaffolding around the AI model: the code that gives it tools, manages its computer and browser, enforces approvals, logs actions, and passes results back. xAI notes Grok 4.7 natively understands the Grok Bot harness, meaning model and scaffolding were designed together. When comparing agents, much of what you feel as quality is the harness, not the model.

What are AI agent approvals?

Tiered permissions controlling which actions run automatically and which pause for a human yes. Typical tiers are automatic for drafts and lookups, notify for routine sends, and require-approval for money, customer contact, and record changes. ChatGPT Dots, Claude Cowork, and Meta Muse all ship approval controls, and any business agent without them should be treated as unfinished.

What is a virtual machine in the context of AI agents?

An isolated cloud computer where the agent works instead of on your laptop: it holds the agent's browser, files, and logins separately from your devices. ChatGPT Dots runs each agent on its own cloud computer, and Meta's Muse Secure VM houses both the agent and the person's data. Isolation limits damage if the agent errs and keeps business activity off personal hardware.

Next step

Pick the three terms from this glossary that appear in the vendor proposal sitting on your desk, and ask the vendor to define each one for your business specifically. Then get a grounded starting point with the free six-step AI automation plan: start your plan. To talk through what you find, book a call.