AI Agent Glossary: 40 Terms Every Marketer Should Know in 2026

This glossary covers the 40 most-used terms in AI agent and multi-agent marketing as of 2026. It is organised into five categories: agent fundamentals, multi-agent architecture, marketing-specific agent terms, technical infrastructure, and emerging concepts. Use it as a quick reference when reading vendor pages, evaluating AI marketing tools, or briefing teams on agent-based workflows. Each term has a one-sentence definition, a longer note where the meaning has shifted from earlier years, and a usage marker showing whether the term is still current, contested, or fading.

TL;DR

• 40 terms across five categories
• Definitions are tightened to current 2026 usage, not earlier definitions
• Where vendors use a term loosely, the contested meaning is flagged
• Emerging concepts include MCP, llms.txt, and world models

How this glossary is organised

The glossary is grouped by where you encounter each term. Fundamentals first, then architecture, then marketing-specific applications, then the technical infrastructure underneath, then the emerging concepts shaping 2026 onwards.

1. Agent fundamentals

Agent. An AI system that takes a goal, reasons about how to achieve it, and uses tools to act. An agent differs from a chatbot in that it pursues an outcome rather than responding to a turn. Used precisely.

Autonomy. The degree to which an agent acts without human approval at each step. Autonomy is a slider, not a binary. Most production agents in 2026 sit between approval-required and fully autonomous.

Tool use. An agent's ability to call external functions or APIs. Tools are how agents do things in the world rather than just produce text. Used precisely.

Reasoning. The agent's process of breaking down a goal into steps. Reasoning quality, not raw model size, is the strongest predictor of agent output quality. Used precisely.

Plan. The structured set of steps an agent intends to take to reach its goal. Plans can be linear or branching. Visible plans are a sign of mature agent platforms.

Loop. The cycle of plan, act, observe, replan that an agent runs while pursuing a goal. Long-running agents iterate this loop dozens of times per task.

Hallucination. Output that is plausible but factually wrong. Mature agent platforms reduce hallucination through grounding (world model), tool use (verified data), and refusal (declining to invent).

Grounding. The practice of constraining an agent's output to verified facts from a known source. Strongly grounded agents produce specific, accurate work. Ungrounded agents produce generic or invented work.

2. Multi-agent architecture

Multi-agent system. A coordinated set of specialist AI agents that share context and collaborate on multi-step tasks. The defining property is shared state, not just multiple prompts. Used precisely.

Orchestrator. The component that routes work across agents, manages dependencies, and handles failures. Without an orchestrator, multiple agents are not really multi-agent. Used precisely.

Specialisation. Designing each agent for one role rather than asking one agent to do everything. The empirical pattern in 2026: specialists outperform generalists on repeated tasks. Used precisely.

Shared state. A common context store that all agents in a system read from and write to. Shared state is the difference between coordinated specialists and isolated prompts. Often called a "world model" in marketing contexts.

World model. A structured knowledge base about a specific business: company, ICP, competitors, voice, brand, performance data. The world model is the shared state for marketing fleets. Becoming standard 2026 vocabulary.

Squad. A group of agents covering one functional area (research, content, growth, etc.). Squads are an organisational layer above individual agents. Vendor-specific term, popularised by Orbitable.

Handoff. When one agent finishes its part of a task and passes the result to another agent. Clean handoffs are the operational signature of a working multi-agent system.

DAG (directed acyclic graph). The structure used by orchestrators to express agent dependencies. DAG-based orchestration is more powerful than linear chains because it supports parallelism. Used precisely.

Parallel execution. Running multiple agents at the same time on independent parts of a task. Parallelism is one of the major advantages of multi-agent systems over single agents. Used precisely.

Concurrency limit. The maximum number of agents the orchestrator will run at once. Tier-aware concurrency (different limits for different customer tiers) is now standard.

3. Marketing-specific agent terms

ICP architect. An agent specialising in defining and scoring ideal customer profiles. The ICP architect typically reads CRM data, win/loss notes, and product usage to produce a scored fit definition.

Buying committee map. An output produced by buyer-mapping agents that lists the people who influence a B2B purchase, scored by role, influence, and engagement. Standard ABM artefact.

Battle card. A sales enablement document that helps reps win against a specific competitor. Modern battle cards are produced and refreshed by competitive-intel agents weekly.

Multi-threading. Engaging multiple stakeholders at a target account in parallel. AI agents accelerate multi-threading by producing role-specific messaging at scale.

Account engagement coverage. The percentage of target accounts where multiple buying committee members have engaged with content. The 2026 ABM metric that matters most.

Net revenue retention (NRR). Quarterly recurring revenue change excluding new logos. AI-driven expansion playbooks reliably add 5 to 10 points of NRR within two quarters.

Pipeline coverage ratio. Open pipeline divided by quota. A typical 3.0x ratio improves to 4.5x or higher with AI-driven ABM motions.

Time to first campaign. Days from world build (or first agent setup) to live execution. Modern fleets target under 7 days.

Plan execution. When the dispatcher agent reads an output, recognises it as a plan, and triggers the named agents in sequence. The mechanism that turns chat into multi-agent execution.

Mission. A coordinated agent run with a defined goal, plan, and progress stream. Missions are the unit of multi-agent work in 2026 platforms.

4. Technical infrastructure

MCP (Model Context Protocol). An emerging standard from Anthropic for connecting AI models to tools and data sources. MCP-compatible agents can be invoked from any MCP-aware client like Claude Desktop or Cursor. Becoming standard 2026 vocabulary.

Streamable HTTP. The transport mechanism MCP uses for long-running agent calls. Replaces older SSE-only patterns. Technical term, used precisely.

API key (mfl format). A platform-specific bearer token used to invoke agents from external systems. Orbitable uses the mfl_live_* prefix.

External API. Programmatic access to a platform's agents from any system. The "Pattern B" pipeline shape lets external tools call individual agents or full playbooks.

Webhook (Pattern A). An outbound HTTP call that fires when something happens inside the agent fleet. Used to integrate with Make.com, Zapier, and similar workflow tools.

Edge middleware. Code that runs at the network edge before requests reach the route handlers. Used in 2026 for enumeration defence, not for auth.

5. Emerging concepts

llms.txt. A simple text file at the root of a website that summarises the site for LLM crawlers. Emerging standard from llmstxt.org. Sites publishing llms.txt see higher AI search citation rates.

Generative engine optimisation (GEO). The practice of optimising content to be cited by generative AI engines like ChatGPT, Perplexity, and Google AI Overviews. Distinct from but adjacent to traditional SEO.

Answer engine optimisation (AEO). Optimising content for direct answer capture, especially in voice and AI assistants. Often used interchangeably with GEO, with AEO leaning toward question-answer format.

Agentic search. Search experiences where an agent does research on the user's behalf, returning a synthesised answer rather than a list of links. Examples include Perplexity Pages and ChatGPT search.

Prompt drift. The slow degradation of an agent's output quality over long sessions or as system prompts accumulate edits. A real production issue, addressed via versioning and evaluation.

Agent evaluation (eval). A repeatable test that scores an agent's output against a fixed rubric. Mature platforms run evals on every prompt change. Borrowed from ML engineering, now central to agent practice.

Common term confusions

Three pairs of terms get conflated. Untangling them helps when reading vendor pages.

When a vendor page uses the bigger term loosely, check what it actually does under the hood.

FAQ

What is the most important AI agent term to understand in 2026?

World model. The presence or absence of a shared world model is the single biggest determinant of whether an AI marketing system produces specific, on-brand output or generic, off-brand output.

What is the difference between AI agents and AI assistants?

Assistants help with conversational tasks turn by turn. Agents pursue goals across multiple steps using tools. The same model can be either, depending on how it is set up.

Are these terms standardised across vendors?

Mostly yes for fundamentals, mostly no for product-specific terms. Squads, missions, and worlds are vendor-specific (Orbitable). Multi-agent, orchestrator, and MCP are industry-wide.

What does "agentic" mean exactly?

Agentic describes systems where AI takes actions toward a goal, not just produces text. Agentic AI is the broader category, AI agents are the specific implementations.

Where can I see these terms used in context?

The Orbitable blog covers most of them in practical posts. The Multi-Agent Marketing Systems Explained post in particular walks through orchestration, shared state, and specialisation with worked examples.

What new terms are emerging that are not in this glossary yet?

Watch for "agent marketplace", "delegated agency", and "agent identity". All three are showing up in 2026 vendor announcements but are not yet stable enough to define precisely.