Agentic AI

A shared vocabulary for the eSAIL team: what it is, who’s conducting, and how to use it well

Gerry Pedraza

Draft: 11/10/26

Part 1

What Are We Even Talking About?

A shared base: what AI and models are, what makes something agentic, and how narrow AI fits in.

From AI to a Model You Can Reach

From AI to a model you can reachArtificial intelligence contains machine learning, which produces specialized models (one function) and general models (many tasks via instructions). Three doors (desktop app, command line, your script) reach the same model. Doors, not categories. Artificial intelligenceSystems doing tasks that need judgment Machine learningLearns patterns from data, not hand-written rules Specialized modelsOne functionCaptions, risk scores, OCR General modelsMany tasks via instructionsOpus, GPT, Gemini A model = the trained result: learned patterns, input to output Desktop app Command line Your script (API) Doors, not categories. The same model can sit behind each door.

Doors to the Model

Doors to the modelFour doors: chat app, assistant inside your tools, command line, your script (API), each with its capabilities. Saved setups (instructions, files, skills) ride on any door; custom GPTs and Gems are becoming skills. Same model; the door decides what it can reach. Chat appSandbox, connectors, projectsEasiest start; vendor runs it Assistant inside your toolsCopilot in Word, Gemini in DocsWorks on the file you have open Command lineWorks on files on your computerRuns commands; for technical users Your script (API)Does exactly what you codeOne-job tool, workflow, or agent Saved setups ride on any doorInstructions + files + skills. Custom GPTs and Gems are becoming skills. Same model. The door decides what it can reach. Then the spectrum test decides what's happening.

Agentic AI: A Working Definition

Agentic AI: working definitionA working definition of agentic AI, its three ingredients (goal, tools, loop), and a spectrum showing who decides the next step: a human in chat, a script in a workflow, or the model in an agent. Anthropic counts workflows and agents as agentic systems. Working definition The AI decides its own next step toward a goal. It acts on things outside the conversation, checks the results, and adjusts, repeating until the goal is met or it needs a human. Three ingredients GoalOutcome, not a prompt ToolsAct outside the chat LoopAct, check, adjust Not a tight definition: a spectrum Test: who decides the next step? A human decidesChat: you are the loop A script decidesWorkflow: fixed path The model decidesAgent: picks its path Anthropic counts both as agentic systems Less agentic More agentic

Agent vs. Agency

Agent vs. agency: the what and the howAgent means acting on someone's behalf, which maps to the what: a goal set by a human that stays fixed along the spectrum. Agency means discretion, which maps to the how: steps, tools, and recovery, which shift from human to script to model. The human sets the what; the agent gets discretion over the how. AgentActs on someone's behalf AgencyDiscretion over choices The whatGoal set by a human The howSteps, tools, recovery Stays fixed along the spectrum Human Script Model Who holds the how shifts The human sets the what. The agent gets discretion over the how.

Narrow AI vs. Agentic AI

Narrow AI vs. agentic AINarrow AI follows one fixed path: lecture audio into a speech-to-text model, captions out. Agentic AI starts from an open goal and loops through plan, act, and check, calling tools such as a captioning model. Narrow AI does one step; agentic AI chooses the steps, and a narrow model can be one of them. Narrow AIOne task, fixed path Input Lecture audio One trained functionSpeech-to-text model Output Captions Never decides what's next Agentic AIOpen goal, chooses its path GoalMake module 3 accessible Plan Act Check Loops until done Tools it can call Captioning model, LMS, files Narrow AI does one step. Agentic AI chooses the steps. And a narrow model can be one of them.

One Agent, Many Kinds of Tools

One agent, many kinds of toolsGoal: an agentic AI podcast with slides. An agent, a general model that researches, writes and decides, calls plain tools (web search, save files, render slides), narrow AI (text-to-speech, speech-to-text) and a general model in one fixed job (an alt-text script). Narrow or agentic describes the job, not the model. Goal: an agentic AI podcast with slides AgentGeneral model: researches, writes, decides Plain toolsNo AI inside Web search Save files Render slides Narrow AISpecialized model Text-to-speech Speech-to-text Narrow useGeneral model, one job Alt-text script Narrow or agentic describes the job, not the model. The same general model can be the agent or a one-job tool.

Part 2

Who’s Conducting?

Workflows, skills, and agents can run the same steps. What changes is who decides the next one.

Anatomy of a Workflow

Anatomy of a workflowA conductor (a script, a no-code flow, or a person with a checklist) runs four steps in a fixed order: extract images (script), describe image (model call), check length (script), save results (plain tool). The workflow is the fixed order, not the scripts. Conductor: follows a fixed pathA script, a no-code flow, or a person with a checklist Extract imagesScript Describe imageModel call Check lengthScript Save resultsPlain tool Steps can be scripts, model calls, tools, or people The workflow is the fixed order, not the scripts. Swap the conductor for the model and it becomes an agent.

Same Steps, Different Conductor

Same steps, different conductorThree columns with the same steps (gather, draft, check). Workflow: a script runs a fixed order, like a machine running a recipe. Guided agent: the model with a skill follows your how but can loop back, like a chef with your recipe card. Open agent: the model alone improvises, like a chef improvising. A skill guides the agent; a workflow replaces it. Workflow Guided agent Open agent Script conductsRuns a fixed order Model + skillFollows your how Model aloneImprovises the how Gather Draft Check Gather Draft Check Gather Draft Check Machine running a recipe Chef with your recipe card Chef improvising Less autonomy More autonomy A skill guides the agent. A workflow replaces it. Same steps. What changes is who conducts.

Someone Always Conducts the Work

Someone always conducts the workFour modes, each defined by who conducts (orchestrates) the work. Chat: you conduct. Workflow: a script conducts a fixed list of steps. Guided agent: the model conducts following a skill. Open agent: the model conducts alone. Someone has to conduct: you, a script, or the model. Conductor = orchestrator: whoever decides the next step ChatYou conductAI answers turnsYou act and checkYou decide next WorkflowA script conductsFixed list of stepsScripts + AI callsSame order always Guided agentModel + skillFollows your howAdapts when neededAsks at checkpoints Open agentModel conductsChooses its stepsLoops until doneNeeds most trust Less AI autonomy More AI autonomy Someone has to conduct: you, a script, or the model. Ask: who is conducting right now?

Part 3

Using It Well and Safely

How to get it, how much to hand over, and how to keep the checking real.

Same Anatomy, Different Names

Same anatomy, different namesProducts such as ChatGPT, Claude, Copilot and Gemini share the same anatomy: a model at the center, connected to a workspace, tools and connectors, persistence, and where it runs. Names change every few months; the anatomy doesn't. ChatGPT Claude Copilot Gemini Same anatomy underneath WorkspaceA sandboxed computer Tools and connectorsReach sites and apps ModelDoes the reasoning PersistenceProjects and files Where it runsCloud or your computer Names change every few months. The anatomy doesn't. Then apply the test: who decides the next step?

Use, Configure, or Build?

Planned visual (V4)

  • Use: turn on agent features in a product you already have. No setup.
  • Configure: no-code. Connect apps, add instructions, schedule tasks.
  • Build: code. SDKs and APIs, custom tools, your own hosting.

Most instructional designers will work at levels 1–2.

The Dial Depends on the Goal

Planned visual (V5)

  • Being agentic is a setting you choose, not a quality score.
  • Turn it right: speed and less effort, but less understanding and control.
  • Turn it left: more control and learning, at the cost of time and attention.

Choose the setting based on the goal, not on what’s possible.

More Control, More Risk, Unless You Shape It

More AI control, more risk, unless you shape itRisk (how far a mistake can travel before a person sees it) rises as the AI conducts more: chat, workflow, guided agent, open agent. Left open, it rises steeply; with guardrails, gently. The gap is your strategy. Levers: clear brief, skill, reference docs, checkpoints, least access. Guardrails lower risk; they never remove it. Risk = how far a mistake can travel before a person sees it Risk Left open: vague goal, broad access With guardrails Your strategy Chat Workflow Guided agent Open agent You conduct A script Model + skill Model alone How much the AI conducts Levers that lower the curve Clear brief Skill Reference docs Checkpoints Least access Guardrails lower risk. They never remove it.

Three Places to Save

Planned visual (V6)

  • Skill: how you like the work done, reusable across projects.
  • Memory: what the assistant should remember about you.
  • Notes: the project’s written state, so work continues between sessions.

Glossary

The basics

Artificial intelligence (AI)
The umbrella term for systems that do tasks that normally need human judgment, such as recognizing speech, predicting outcomes, or writing text.
Machine learning
How most modern AI is made: the system learns patterns from examples instead of following rules someone wrote. A rules-based accessibility checker is not machine learning; nobody trained it.
Model
What training produces: learned patterns that turn an input into an output. When people say “the AI,” they usually mean the model.
Specialized model
A model trained for one function, such as captions, text-to-speech, OCR, or a risk score. It cannot do anything else.
General model (LLM, foundation model)
A model that learned language broadly, such as Opus, GPT, or Gemini. It can do many tasks depending on the instructions it gets.

Doors and setups

Door (interface)
How you reach a model: a chat app, an assistant inside your tools, a command line, or your own script. The same model can sit behind each door; the door decides what it can reach.
API
The way a script talks to a model directly. A script can do exactly what it is coded to do and nothing more.
Connector
A link that lets an AI app reach another system, such as your email, files, or LMS.
Workspace (sandbox)
A computer the AI can work in, separate from yours, where it can run code and create files.
Saved setup
Instructions plus files (plus skills) saved for reuse. Custom GPTs and Gems were saved setups; vendors are moving them into skills.
Skill
Written guidance, sometimes with scripts or files attached, that the model reads and follows. The model still conducts and can adapt. A skill guides the agent; a workflow replaces it.

Who decides the next step

Narrow AI
In this talk, reserved for specialized models: one model, one function. “Narrow” describes scope, not quality.
One-job tool
Anything that does one fixed step: a specialized model, a general model locked to one instruction (such as an alt-text script), or a plain tool.
Workflow
Several steps run in an order fixed in advance by a script, a no-code flow, or a person with a checklist. The model may do steps, but it never chooses the path.
Agent
A system where the model decides its own next step toward a goal, acts on things outside the conversation, checks the results, and adjusts.
Agentic system
Anthropic’s umbrella term covering both workflows and agents.
Agency
Discretion over how to act. The human sets the what; the agent gets discretion over the how.
Conductor (orchestrator)
Whoever decides the next step: you in chat, a script in a workflow, or the model in an agent. “Orchestrator” is the industry term.
Guided agent
A model that conducts while following a skill: it uses your preferred way of working but can loop back or adjust.
Open agent
A model that conducts with no skill or set process, improvising its steps.

Working safely

Delegation brief
A request written like a hand-off to a colleague: outcome, materials, what done looks like, boundaries, and checkpoints.
Plan mode
A permission checkpoint: the agent may read and research but takes no action until you approve its plan. You can always ask for it in your brief.
Checkpoint
A planned point where the agent stops and waits for a person to review before continuing.
Least access
Giving an agent only the files, apps, and permissions the task needs.
Rubber-stamping
Approving an agent’s work by reflex after it has been right several times. A human is in the loop on paper but not in practice.
Prompt injection
Hidden instructions inside a web page, email, or file that try to steer an agent that reads them.