The Foundation

Agentic AI—An AI system that can plan and execute multistep tasks autonomously toward a defined goal, taking action on systems and data with minimal human intervention. The defining shift from generative AI (chatbots that answer questions) to agentic AI is the move from "tell me something" to "do something."

AI Agent—A specific software entity built around a language model that perceives its environment, reasons about goals, and acts using tools (APIs, databases, other agents). Think of an agent as a digital worker with a defined scope, set of tools, and authority level.

Frontier Models—The most capable AI models at the cutting edge of development (Claude, GPT, Gemini, and similar). They are powerful but expensive and change rapidly enough that any specific model choice is obsolete within months. Most enterprises consume these via API rather than building them.

Probabilistic (vs. Deterministic)—A defining characteristic of AI systems: the same input can produce different outputs depending on context, memory, and model state. Unlike traditional software, which gives the same answer every time, AI is variable. This variability has major implications for testing, governance, and risk management.

Stateful (vs. Stateless)—Agents maintain memory, context, and reasoning state across interactions rather than treating each interaction as independent. State persistence enables long-running workflows but also creates the risk that agent behavior drifts over time as the underlying data and models change.

Compounding—The defining property of AI advantage that did not exist in prior technology waves: Data improves agents, agents improve people, people redesign work, and the redesigned work generates better data. The flywheel turns on its own. The gap between AI leaders and followers gets structurally harder to close every quarter, which is why early movers pull away faster than in any previous technology wave.

The Infrastructure

Enterprise Orchestration—The layer where a company manages its agents, its encoded know-how (skills, tools, and data ontology), and their access to data and systems as a single, governed enterprise asset rather than a scattered collection of vendor-supplied tools. It spans the registries that record what exists, the boundary at which agents access data and systems, and the controls that keep the estate inspectable and changeable. Distinct from agent orchestration, it is the narrower task of coordinating which agent does what within a single workflow.

Harness—The execution infrastructure wrapped around an AI model that turns a probabilistic language model into a reliable working agent. Includes tool integration, memory management, safety controls, and observability. In 2026, the harness became the competitive frontier: The model is the engine, but the harness is the vehicle. Agent = Model + Harness.

Multi-Agent System—A configuration where multiple AI agents work together on a workflow, each handling a specific subtask and passing results downstream. It is common in complex agentic work (e.g., a research agent feeding a writing agent, which then feeds a fact-checking agent). Multi-agent systems are where cascade risk lives.

Model Context Protocol (MCP)—An open standard that defines how AI agents connect to external tools, data sources, and other systems in a consistent, governable way. Now stewarded by the Linux Foundation under the Agentic AI Foundation umbrella. Think of MCP as the USB-C of agentic AI: a common interface that lets any agent plug into any tool without bespoke integration work for each combination.

Semantic Layer—A shared business vocabulary that defines, centrally and consistently, what your key business terms mean—revenue, customer, churn, margin, product—along with the rules and relationships between them. It sits between raw data and AI agents, so every agent reasons from the same definitions. Without it, every agent invents its own dialect of your business.

AI Estate—The full inventory of AI agents, models, tools, integrations, and data pipelines operating across the enterprise. Managing the AI estate is the new infrastructure discipline, analogous to managing the IT estate or real estate portfolio. If you cannot inventory it, you cannot govern it.

Shared Memory Layer—A persistent knowledge store that agents can read from and write to, allowing learning to compound across agents and over time. Without it, every new agent starts from scratch. With it, the tenth agent is faster and smarter on day one than the third agent was after six months.

The Controls

Registries (Agent/Tool/Skill)—Central records of every agent, tool, and skill operating in the enterprise, including purpose, owner, capabilities, and access permissions. If something is not registered, it does not run. The registry is the control surface that enables governance.

Governed Gateway—The infrastructure layer through which every agent tool call routes, enabling policy enforcement, telemetry, and audit trails. Governance happens at this boundary, not inside the agent’s reasoning, which is technically inaccessible to inspection.

Promotion Gates—Automated quality checks that determine whether an agent can move from development to production. Functions like quality control on a manufacturing line: the gate passes or blocks automatically, based on test results, threat assessments, and policy compliance.

Evals (Evaluations)—The structured, continuous testing of AI agents and models to verify they produce accurate, safe, and useful outputs. Unlike traditional software testing, evals run continuously because AI behavior can drift as data shifts and underlying models update. Industry rule of thumb: Building the agent is roughly 20% of the work; evals are 60%; ongoing monitoring is the remaining 20% and never stops.

Golden Test Suites—A defined set of test cases representing known scenarios that agents must pass before reaching production. Like a final exam, an agent has to retake it every time it or its underlying model changes.

Shadow Mode—Running an agent in parallel with existing systems or human decision-makers without acting on its outputs. It is used to validate behavior before giving it real authority. Helps surface edge cases without putting business operations at risk.

The Costs

Tokens/Token Costs—The unit of compute that AI models consume to process and generate text. Every agent action consumes tokens, and at enterprise scale, aggregate token cost becomes a material line item that is invisible to traditional IT cost reporting. Token cost reporting is to AI operations what unit economics are to a manufacturing line.

AI-Native Software Development—A development model where AI agents handle code generation, testing, and documentation while human engineers handle architecture and direction, achieving five to ten times the productivity of traditional development. The function most likely to transform first into an AI-led organization, and the clearest proof point of what an AI-transformed operating model looks like.

The Risks

Prompt Injection—An attack where malicious instructions are inserted into the input an AI agent processes, attempting to override the agent’s intended behavior. Analogous to social engineering against humans, but at machine scale and machine speed.

Cascade Risk—The risk that an error or compromise in one agent propagates through downstream agents in a multi-agent chain, triggering a series of autonomous actions before anyone realizes something is wrong. One of the most underappreciated operational risks in enterprise agentic AI.

The People and Governance

Tinkerers—Employees outside traditional technology roles who naturally experiment with AI tools and discover use cases that no central team would have designed. A leader’s tinkerers are the organization’s sensing network for what’s actually working. Identifying and equipping them is a strategic move, not a culture program.

Two Governance Motions (Run the Business/Change the Business)—The recognition that AI transformation requires a change-the-business governance motion (focused on learning, pivots, multiyear outcomes) running in parallel with existing run-the-business governance (focused on reliability, efficiency, milestone delivery). Running AI through existing governance alone is the single most common reason that transformations stall.