Agentic AI Glossary.
Plain language for
business leaders.
Cut through the noise by knowing the terms. Walk into every AI conversation confidently.
Part of the SERAA Cortex, Agent Design Studio is where the objectives and performance parameters of an AI agent are defined and fine-tuned. It gives your team precise control over what an agent does, how it behaves, and what success looks like, before the agent is ever deployed in a live environment.
A dedicated portal within SERAA Cortex where AI agents, once built, tested, and approved, are made available for use. Think of it as an internal or external app store for AI agents. Business teams can browse, select, and deploy agents directly from the marketplace, without needing to involve technical teams for every new use case.
The system of record for every AI agent an organization runs: what each one is, who owns it, and what it is permitted to reach. A registry's authority is set by what happens to an agent that is not in it.
The operational discipline for running AI agents in production, covering build, evaluation, release, observation and governance as one continuous loop. It is the first Ops practice with no release boundary, because four of the five things that determine an agent's behaviour change without anyone deploying anything.
What happens when an organization's AI agents multiply faster than its ability to see, govern and account for them. It is a rate problem rather than a headcount problem: four hundred governed agents is an estate, and twelve ungoverned ones is sprawl.
The libraries you build agents with, providing the loop, the state model, and the handoffs between agents. A framework can only govern agents that import it, which is why the choice matters less than teams expect and never solves the estate-level problem.
How an agent proves what it is and on whose behalf it is acting, at the moment it acts. Identity systems have two categories, humans and services, and an agent is neither, because the reach of its credential cannot be known when you issue it.
The build environment within SERAA Cortex where AI agents are created. Agent Studio provides the tools and structure needed to design agents that can reason, take action, and deliver outcomes, without requiring your team to start from scratch each time.
Before an AI agent goes live, it needs to be tested rigorously. AgentEval, also referred to as the Agent Test Bench, is the testing environment within SERAA Cortex where agents are evaluated for accuracy, reliability, and performance. Only agents that meet the required standards move forward to deployment.
The category of systems that pursue goals by choosing their own actions, rather than producing an output for a person to act on. Every earlier wave of AI kept a human between the decision and the consequence. This one removes them by design.
Agentic RAG hands the retrieval loop to the agent, so it decides whether to search, what to search for, whether the result is good enough, and whether to go again. The shift is one of discretion rather than accuracy, and the hard part is the stopping condition: knowing when to answer rather than search once more.
An agentic workflow is a business process where some steps are decided at runtime by a model instead of being specified in advance. It is a spectrum, not a category. What works in production is usually a deterministic process with one or two agentic steps inside it, and the useful question is which steps those are.
A software system that pursues a goal by deciding its own next step, calling tools and other systems to act, then observing the result before deciding again. What separates an agent from automation is discretion: nobody wrote the sequence in advance.
AI TRiSM bundles the controls that keep AI explainable, governed, secure and privacy-compliant into one name, because those concerns used to sit with four teams and no budget line covered the whole. It is a market category and not a standard. Nothing certifies against it, so nobody is AI TRiSM compliant.
Capturing what an agent did, step by step, as traces you can query and replay. It exists because agents fail in ways that look like success: every conventional monitoring signal stays green while the agent competently does the wrong thing.
The single operational surface from which an organisation monitors every agent in production and intervenes in it: live visibility, cost per agent, policy on each action, an audit record, and the ability to stop an agent immediately. Observability tells you what happened; a control tower is where you act on it.
The open standard for agents discovering each other and delegating tasks across platforms. Governed by the Linux Foundation, stable since 2026, and it treats the remote agent as a black box. Which makes A2A the first protocol that lets you delegate work you cannot observe.
AI bias is a repeatable pattern of unfair outcomes rather than a one-off error. Three different things share the name: statistical skew, discriminatory impact, and evaluator preference. Only the second is what regulators mean. It also gets harder when an agent acts, because every individual step can pass while the sequence does not.
The decision rights, controls, and evidence that determine what an organization's AI systems may do, who answers for them, and how that can be proved. Policy states the intent; governance is the part that constrains what actually happens.
An AI platform is the shared layer your AI runs on. Four unrelated product categories are sold under that one name: infrastructure, model serving, model lifecycle and agent operations. Most shortlists mix two or three of them, which is why the evaluation stalls and the market looks immature when it is the shortlist that is mixed.
Controls that keep an AI system inside its permitted behaviour, applied outside the model rather than written into its instructions. The test that separates a guardrail from a wish: ask what happens if the model ignores it.
Fluent, confident, false output. Not a malfunction but the same mechanism that produces correct answers, since the model generates the most plausible continuation and plausibility comes apart from truth. In an agent the worst case is not a wrong fact. It is the agent reporting an action it never took.
The coordination layer that decides which AI models, agents, and tools act on a task, in what order, and what happens when a step fails. It holds the state, the routing, and the accountability that no individual component carries alone.
An AI Center of Excellence is a central team holding the platform, standards and shared assets so every business unit is not solving the same problems separately. The hard question is not what it does but how much it centralises, and the answer changes as you grow: the model that fits three use cases is the bottleneck at thirty.
A record complete and defensible enough for a third party to reconstruct a specific decision months later and establish on whose authority it was taken. The hardest requirement is not retention. It is attribution, and most deployments cannot name the human behind a given action.
AI that operates independently to complete tasks without requiring human input at each step. Autonomous AI is capable of managing its own decision-making within defined boundaries. The degree of autonomy ranges from supervised systems that seek approval before acting, to fully autonomous systems that operate independently within established guardrails.
Establishing whether an AI agent is doing its job, judged on the path it takes as well as the answer it produces. An agent that reaches the right outcome through a dangerous or wasteful sequence of tool calls has still failed.
Relabelling a product as agentic AI without rebuilding it: chatbots, assistants and scripted automation renamed rather than re-engineered. Gartner named the practice in June 2025. Because "agent" has no agreed boundary, vendors set the threshold wherever their product happens to sit.
The ability of agents built by different teams, vendors and frameworks to discover each other, exchange work and produce a joint result. It runs across four layers, and only the wire protocol has a settled standard. Identity, shared meaning and accountability are still yours to solve.
AI governance is the set of decision rights, controls and evidence behind what your AI is allowed to do and who answers for it. Policy states the intention; governance is what makes the intention bite. Agentic systems change the shape of the problem, because an approval given once now covers behaviour that keeps changing.
Agent memory is what an agent keeps after a run ends and can pull back in later. That differs from the context window, which empties. It matters most as a security question: an injected instruction in a context window pays off once, and the same instruction written to memory pays off on every future retrieval.
An agent that diagnoses and fixes without waiting on anyone to sign off each step. The question of whether compliance permits this treats autonomy as something the whole function either has or lacks. It attaches to individual actions instead. Restart a stateless pod, yes. Migrate a schema, no.
The adherence of an AI system to applicable laws, regulations, and internal policies. In regulated industries, non-compliance by an AI system carries the same legal and financial consequences as non-compliance by a human team. AI compliance requires that agents operate within defined boundaries, produce auditable outputs, and can demonstrate the basis for their decisions when asked.
An enterprise agentic AI platform company. Covasant SERAA, a full-stack agentic AI platform that enables organizations to build AI agents, orchestrate them across workflows, and govern them at scale. Covasant's positioning is as a platform company: organizations can build agents on SERAA Cortex themselves, or Covasant can build for them. The enterprise products already deployed on SERAA Cortex are proof that it works.
A context graph records not just what is asserted but the conditions under which it holds: when it became true, where it applies, which source said so, and who may rely on it. A knowledge graph tells an agent what is true. Those four properties are what make a fact safe to act on rather than merely readable.
The process of collecting and importing data from various sources into a central system. Within Covasant's Enterprise Data Management System, Data Ingestion is the first stage, connecting to structured and unstructured data sources across your organization through 200-plus enterprise connectors, making that data available for processing and AI-driven analysis.
The capability that makes enterprise data understandable to systems rather than only to people: what each field means, how entities relate, how current it is, and who may use it. Business intelligence made data readable by analysts; this makes it usable by machines.
A data lakehouse puts warehouse guarantees, meaning transactions, schema enforcement and governance, onto low-cost object storage. It solved the storage problem and left the one agents care about. Time travel, lineage and row-level permissions are already in there, and almost nobody wires any of it to their agents.
The stage between data collection and data use, where raw data is cleaned, corrected, and made ready for analysis. In Covasant's platform, Data Processing is where our MDM solution operates, ensuring that the data flowing into AI agents meets the quality standards that make their outputs reliable. Garbage in, garbage out. Data Processing prevents garbage from ever entering.
Data quality for AI is fitness for a consumer that cannot sanity-check. Every quality programme ever built assumed somebody at the end who would notice a number looked wrong and query it. Remove that person and a dataset good enough for reporting becomes a system taking a wrong action every fiftieth run.
The binding constraint on enterprise AI has never been how capable the model is. It was getting to the data, then wiring into the systems, and now proving who answers for what a system did. Because the term predates agents, much of the guidance still circulating solves a problem you no longer have.
A connector is a pre-built integration that lets a platform read a source system without bespoke engineering. Every vendor advertises a count, and the number tells you almost nothing: six connectors that handle your systems properly beat two hundred that nearly do. Ask what happens when a record is deleted.
Enterprise data management is how an organisation keeps its data available, correct and governed. It was built for consumers you could name in advance, and agents are an unknown number of unknown consumers asking unanticipated questions. You do not have to finish it before running agents, but you do need to know which parts are load-bearing.
Three categories of financial and operational risk that cost organizations, and governments, significant resources each year. Fraud involves intentional deception for financial gain. Waste refers to the misuse or overuse of resources without resulting benefit. Abuse involves practices that are not necessarily illegal but deviate from accepted standards. Agentic AI can detect patterns across all three categories at a scale and speed that human audit teams cannot match.
Generative AI produces new output rather than retrieving or scoring something that exists. Because the output space has no edges, you cannot certify behaviour before deployment the way predictive models allowed, so assurance moved to runtime. It is also not an alternative to agentic AI; it is the component agents reason with.
GraphRAG is retrieval that uses graph structure rather than vector similarity alone. It is sold on multi-hop questions, and the paper that introduced it claims something narrower: answering questions about a whole corpus at once, which retrieval cannot do because retrieval returns a subset. The cost sits at indexing.
GreenOps puts carbon and energy alongside cost in the metrics an operations team already watches, which works because idle capacity is visible and switchable. Agentic workloads break that. An agent consumes per action, decides its own number of actions, and leaves nothing idle for anyone to find.
See AI Governance. The policies, processes, and controls that govern how AI is built, deployed, monitored, and retired within an organization. Without governance, AI systems can drift from their intended purpose, violate regulatory requirements, or cause harm without accountability. Governance is not a constraint on AI capability, it is the structure that makes AI trustworthy enough to deploy at enterprise scale.
See AI Guardrails. Boundaries embedded into an AI system to ensure it only acts within approved parameters. In Covasant SERAA guardrails are enforced structurally, through the Agent Registry, which requires every agent to be documented and approved before deployment, and through the AI Agent Control Tower, which monitors behaviour in real time and provides the mechanism to stop any agent that acts outside its defined scope.
An AI agent kill switch stops a running agent, and the hard part is not the button. The EU AI Act asks for a halt in a safe state, which means deciding what happens to work already in flight, and whether stopping one agent also stops everything it has already delegated.
A knowledge graph records entities and the relationships between them in a form a machine can walk, with an agreed schema above it saying what kinds of thing exist and how they may relate. Projects fail on that agreement, not on the technology. Deciding what a customer is turns out to be a political negotiation in engineering clothes.
A discipline that ensures an organization has a single, consistent, and accurate version of its critical data, such as customer records, product information, supplier lists, and financial hierarchies. Without MDM, the same entity may appear with different names, IDs, or attributes across different systems, causing inconsistency and errors throughout any process that depends on that data, including AI-driven processes.
The coordination of multiple AI agents working together, often simultaneously, to complete complex tasks or workflows. Multi-Agent Orchestration in Covasant SERAA enables agents with different specializations to collaborate: one agent might gather data, another analyse it, a third draft a report, and a fourth route it to the appropriate decision-maker, all without human coordination at each handoff. This is the architecture that makes enterprise-scale agentic AI practically deployable.
A network of AI agents, each with its own role and capabilities, that work together to achieve broader objectives. Multi-agent systems are the architecture underlying enterprise agentic AI deployments. Rather than one general-purpose agent doing everything, a multi-agent system deploys specialised agents that each do one thing well, and the orchestration layer coordinates them into a coherent, governed whole.
Model routing sends each request to the model that fits it, instead of paying frontier prices for questions a small model answers correctly. The saving is real, and the two routing errors are not equal: routing up costs money, routing down costs correctness. Which is why a record of what served each request matters.
MCP is an open standard connecting AI applications to tools and data through one client-server interface, covering tools the model can call, resources it can read and prompt templates it can use. It won by turning an integration problem into a single interface. It also moved the trust decision from build time to runtime.
A way of interacting with a software system using everyday spoken or written language, rather than technical commands or structured queries. In SERAA Cortex, the AI Agent Control Tower uses a natural language interface so that business users and engineers can query agent performance, ask questions about their workflows, and receive clear answers without writing code or learning a new technical system.
The ability of an AI system to understand, interpret, and generate human language. NLP is foundational to how AI agents read unstructured documents, extract relevant information from text-heavy sources, and produce readable outputs. For organizations dealing with large volumes of contracts, correspondence, reports, and regulatory filings, NLP is what makes it possible for AI to work with those documents the same way a human analyst would, only faster and at greater scale.
A technique used in AI systems where the model retrieves relevant information from an external data source before generating a response. Rather than relying solely on what was baked into the model during training, RAG allows AI to reference live, specific, proprietary data in real time, producing outputs that are grounded in your actual enterprise information. In the context of agentic AI, RAG is a key mechanism for giving agents access to the current state of your organization's knowledge and data.
Robotic process automation drives software the way a person would, clicking and typing rather than integrating. That is why it scaled, and why it breaks: it depends on a screen somebody else controls. Agents do not remove the dependency, they trade a loud failure for a quiet one.
Data that is organized in a defined format, typically in tables, rows, and columns, as found in spreadsheets, databases, and ERP systems. Structured data is easier for traditional analytics tools to process. However, it represents only a fraction of the data most organizations generate. The rest is unstructured. SERAA Axon, Data Intelligence layer is designed to work with both, ensuring AI agents have access to the full picture.
Shadow AI means AI running inside the business that nobody accountable knows about. Shadow agents add the ability to act rather than only answer, which is a different exposure entirely. Shadow IT was closed by revoking access, and a disclosure cannot be revoked once data has reached somebody else's model.
A semantic layer holds your metrics, dimensions and joins once and serves the same definitions to everybody. It used to exist so two dashboards would agree. It now exists because an agent pointed at raw tables re-derives your business logic on every prompt, and it cannot ask a colleague what revenue means here.
Tool calling is how a language model gets to use software, and the name describes something the model never does. It emits text naming a tool and some arguments. Your code decides whether to run it, which means every tool call that happens was a decision your software made.
The process of identifying, assessing, and managing the risks that arise from working with external vendors, suppliers, and partners. Sixty percent of enterprise data breaches originate from third parties. Covasant's TPRM platform, replaces periodic manual assessments with continuous AI agent monitoring, providing dynamic risk scoring, automated governance, and real-time alerts across your entire vendor ecosystem. It is the difference between knowing your vendor risk picture today versus when your next scheduled review arrives.
Data that does not follow a predefined format, such as emails, contracts, audio recordings, images, and documents. Unstructured data accounts for the majority of information generated by most organizations, yet traditional analytics tools struggle to process it. Agentic AI, through SERAA Axon, Data Intelligence layer, can read, interpret, and act on unstructured data the same way a skilled analyst would, making the full breadth of your enterprise information available as a source of intelligence, not just the portion already formatted for a database.
The use of technology to carry out a sequence of business tasks with minimal human involvement. Agentic AI takes workflow automation significantly further, allowing systems to follow predefined steps, reason through decisions, adapt to new information, and complete complex processes end to end. Where traditional workflow automation breaks when it encounters an exception, an agentic AI workflow reasons about the exception and resolves it, escalating to a human only when genuine judgment is required.
Proof that the platform works. Every product built on Covasant SERAA
Every Covasant enterprise product is built using Covasant SERAA, the same platform available to your organization. These products exist as proof that SERAA Cortex delivers enterprise-grade agentic AI at scale, across the most demanding compliance, security, and operational environments.
Ready to go beyond the glossary
and see Agentic AI in action?
Speak to a Covasant AI specialist. See a live product demo of Covasant SERAA. Or ask us the questions that the glossary raised but did not fully answer.