Agentic AI Explained

5.1  The definition

Lesson 3 described a system that answers. This lesson describes a system that does.

An agent is a model that can perceive a situation, plan, take actions through tools, observe the results of those actions, and then decide what to do next — repeating until the goal is met or it gives up. The standard component list is perception, reasoning, memory, planning, and tool-based action such as web search, code execution, API calls, and file access.1

Strip the vocabulary away and three things are being added to the chat model you already understand:

1. Tools — functions the model can call, whose results come back into its context.
2. A loop — the output of one step becomes the input to the next, without a human pressing Enter each time.
3. A stopping condition — something that decides when to halt.

Hold on to that third item. It is the one that is hardest to get right and the one most often left out of the marketing.

FIG-05-01

Model, tools, loop

A single clear cycle diagram: Goal enters at the left; a loop of four stations — Plan → Act (tool call) → Observe → Decide — with the decision branching either back into the loop or out to Done. Draw the tool call as reaching out of the loop to a rack of labelled tools (search, files, code, API) and bringing something back. Mark two points on the diagram in the caution colour: the arrow returning into the loop, annotated errors compound here, and the Decide station, annotated who decides "done"?

Aspect 16:9 · 3200×1800 px · Brand palette: navy #00040c, silver-blue #a8c0d8, column gold #a87848, caution #8f3a2a · Style: technical cycle diagram

5.2  Tools and the protocol layer

Tools raise an unglamorous problem that turns out to be the story of 2025 and 2026. If every AI application has to build a custom integration with every tool it might use, the work required grows as the product of the two counts. Ten applications and ten tools is a hundred integrations, each separately built and separately maintained.

Anthropic released the Model Context Protocol in November 2024 as an open standard for connecting AI systems to external tools, data sources, and applications, defining a client–server interface through which an agent reaches outside capabilities.2 The useful analogy is USB-C: replacing many bespoke connectors with one specification, so that N×M custom integrations collapse to N+M.3

In December 2025 the governance question was settled in public. On 9 December the Linux Foundation announced the formation of the Agentic AI Foundation, anchored by three donated projects: Anthropic's Model Context Protocol, Block's goose, and OpenAI's AGENTS.md. The foundation was co-founded by Anthropic, Block, and OpenAI, with support from Google, Microsoft, AWS, Cloudflare, and Bloomberg.24 A protocol owned by one vendor became a protocol owned by a neutral foundation — which is normally the moment a standard stops being a strategy and starts being infrastructure.

Adoption pressure is real regardless of which count you prefer. Gartner forecasts that 40% of enterprise applications will be integrated with task-specific AI agents by the end of 2026, up from under 5% in 2025.6 Treat that as a well-sourced projection rather than a measurement — it is a forecast, and forecasts about enterprise adoption have a poor historical record.

5.3  Where agents break

Three failure modes. They are structural, not teething problems, and each has a countermeasure.

1. Compounding error

Lesson 3 established that each step is a prediction. In a chat, you see every prediction and correct it. In a loop, step two consumes step one's output unexamined. If each step is 95% reliable, ten steps is not 95% reliable — it is roughly 60%. Long autonomous chains are the worst possible structure for a component that is usually right.

Countermeasure: shorten the chain, and put a verification step inside the loop rather than only at the end. Tasks with cheap automatic checks — code that either compiles or does not — degrade far more gracefully than tasks where errors are silent.

2. No natural stopping condition

A chat ends when you stop typing. A loop has to decide, from the inside, that the goal is met — and “I have done enough” is exactly the judgement Lesson 6 will show these systems are trained not to make. Agents that cannot tell they are finished either stop too early with a confident summary of incomplete work, or keep going, consuming budget.

Countermeasure: impose the stopping condition from outside. Step limits, budget limits, wall-clock limits, and explicit success criteria written before the run starts.

3. Unbounded permissions

The failure with the largest downside. A tool-using agent acts with whatever authority you gave it, and it will use that authority on the basis of a prediction. An agent with write access to a production system, a payment method, or an outbound mailbox can do real and irreversible damage without any of the steps being individually unreasonable.

Countermeasure: least privilege, and a human confirmation gate on anything that is outward-facing, costly, or hard to reverse. This is the same instinct Lesson 9 applies to sovereignty at national scale — decide deliberately what you are handing over.

5.4  Do you need an agent at all?

The most useful guidance in this area comes from the people building the tools, and it is deflationary. Anthropic's engineering guidance distinguishes workflows — systems where models and tools are orchestrated through predefined code paths — from agents, where the model dynamically directs its own process and tool use. Its explicit recommendation is to find the simplest solution possible and increase complexity only when needed, which may mean not building an agentic system at all; the most successful implementations observed used simple, composable patterns rather than frameworks.7

A practical decision rule follows. Use a workflow when you know the steps in advance — it is predictable, cheaper, and debuggable. Use an agent when the steps genuinely cannot be known ahead of time and the flexibility is worth the loss of predictability. Most tasks people reach for an agent to solve are workflows wearing a costume.

Demonstration protocol

1. Give an agent a three-step task with a deliberate obstacle in step two — a file that is not where the instructions say, or a source that returns an error.
2. Show it noticing and adapting. This is the genuine capability and it is worth showing honestly.
3. Show where you intervened. Non-negotiable. Every real agent run includes a moment where a human redirected it, and cutting that moment is the most common dishonesty in AI demonstrations.
4. Show a permission gate firing. Set up the run so the agent must ask before doing something irreversible. Let the viewer see the confirmation prompt.
5. Optional, and the most instructive minute available: run the same task as a fixed workflow. If the workflow wins, say so on camera.