Agent
An AI setup that can pursue a task across steps.
- What it does
- Plans, uses tools, checks results, and keeps going until a goal or stopping rule is met.
- Why it matters
- Useful when the work is more than one prompt, such as research, cleanup, coding, or repeated follow-up.
- Use it for
- Tasks with clear rules, visible progress, and a human review point.
- Tools to know
- Start with an approved assistant or automation tool. Move to agent builders only when the steps are clear enough to supervise.
- Watch out
- Do not hand it vague authority. Define what it can change, what it can only draft, and when it must stop.
Agent harness
The rails around an agent.
- What it does
- Provides tools, memory, approvals, logs, tests, and limits around the agent.
- Why it matters
- The harness is often what makes an agent useful or risky. The model is only one part.
- Use it for
- Work where reliability matters, like content production, support triage, data cleanup, or internal operations.
- Tools to know
- Look for systems with approval steps, activity logs, tool permissions, and a clear place to review the output.
- Watch out
- An agent without approvals and logs is hard to trust after something goes wrong.
Automation
A repeated task that runs with less manual effort.
- What it does
- Moves a known task from manual steps into a repeatable system.
- Why it matters
- AI can help map the steps before you connect tools.
- Use it for
- Tasks with a clear trigger, inputs, owner, output, and failure case.
- Tools to know
- Zapier, Make, Microsoft Power Automate, and built-in automations inside your CRM, email, or project tools.
- Watch out
- Do not automate a messy process before you understand the judgment points.
Context window
The working memory the model can reference.
- What it does
- Holds the instructions, chat history, documents, tool results, and output space the model can use in one request or conversation.
- Why it matters
- More room helps with longer work, but too much clutter can make answers worse.
- Use it for
- Long documents, multi-step analysis, and projects where earlier details matter.
- Tools to know
- Use assistants with long-document support when the task needs it. Use shorter, cleaner chats for quick work.
- Watch out
- Clean context beats huge context. Remove old noise, stale drafts, and irrelevant tool results.
Token limit
The budget for what fits in the model's working space.
- What it does
- Tokens are chunks of text. Inputs, outputs, tools, and sometimes reasoning all spend the budget.
- Why it matters
- When the budget runs out, the model may stop, drop older context, or need a shorter prompt.
- Use it for
- Planning long tasks, big document work, and agent sessions that need summaries.
- Tools to know
- Use document summaries, project briefs, and saved prompt cards to keep long work from becoming one overloaded chat.
- Watch out
- If a chat gets confused, do not only blame the model. The context may be crowded.
MCP
A common way for AI apps to connect to tools and data.
- What it does
- Model Context Protocol lets AI applications connect to external systems such as files, databases, search tools, and business apps.
- Why it matters
- It can reduce one-off integrations and make tools available across AI clients that support the protocol.
- Use it for
- Connecting approved business data and tools to an AI assistant with clear permissions.
- Tools to know
- Think in terms of MCP clients and MCP servers. Your team should know who owns the server, permissions, and data access.
- Watch out
- Permissions, authentication, and tool trust matter. Treat an MCP server like a real software integration, not a harmless plugin.
Prompt
The task instructions you give the model.
- What it does
- Tells the model what role to play, what context to use, what output to create, and what constraints to follow.
- Why it matters
- A better prompt makes the work easier to judge because the output has a target.
- Use it for
- Briefs, rewrites, questions, comparisons, checklists, and follow-ups.
- Tools to know
- Use reusable prompt cards, team examples, and approved assistants. The best prompt usually starts from a real work artifact.
- Watch out
- Do not paste private details just to make the prompt feel complete. Use a safe summary first.
Model
The AI engine doing the prediction or reasoning.
- What it does
- Generates text, code, images, structured data, analysis, or tool requests based on input.
- Why it matters
- Different models are better at different jobs: writing, reasoning, speed, cost, coding, long context, or multimodal work.
- Use it for
- Match the model to the work quality needed, not the fanciest name available.
- Tools to know
- Compare the models available inside your approved tools first. A paid business assistant may matter more than a new model name.
- Watch out
- A stronger model will not fix unclear instructions, bad context, or missing review.
Inference
The moment a model generates an answer.
- What it does
- Runs your input through the model and produces output, often with settings for speed, cost, length, and randomness.
- Why it matters
- This is where usage cost, latency, and output quality show up.
- Use it for
- Comparing whether a task needs a quick draft, careful reasoning, or a structured output.
- Tools to know
- Most business users see this through the assistant interface. Builders see it through APIs, usage dashboards, and model settings.
- Watch out
- If the same task gives uneven answers, check prompt clarity, model settings, and the examples provided.
RAG
Retrieval-augmented generation.
- What it does
- Finds relevant source material first, then gives that material to the model before it answers.
- Why it matters
- Helps answer from your documents or current sources instead of relying only on what the model already knows.
- Use it for
- Policies, knowledge bases, past reports, product docs, customer help, and research notes.
- Tools to know
- Look for knowledge-base search, file-connected assistants, enterprise search, or internal chat tools connected to approved sources.
- Watch out
- Bad retrieval creates bad answers. Test whether it found the right sources before judging the model.
Embeddings
Numbers that help computers compare meaning.
- What it does
- Turns text into vectors so related text can be found, grouped, recommended, or classified.
- Why it matters
- This is one common backbone for semantic search and retrieval systems.
- Use it for
- Search across notes, support articles, docs, saved examples, or content archives.
- Tools to know
- Most teams meet embeddings through semantic search, vector databases, knowledge bases, and document-retrieval features.
- Watch out
- Embeddings help find related material. They do not decide whether the material is correct.
Eval
A test for whether the AI output is good enough.
- What it does
- Checks outputs against expected behavior, examples, rubrics, or human review.
- Why it matters
- If repeated work matters, you need a way to know whether it improved or drifted.
- Use it for
- Prompt changes, agent tasks, customer-facing answers, extraction, and high-volume content checks.
- Tools to know
- Use sample sets, checklists, review rubrics, and logs before moving to formal eval platforms.
- Watch out
- A score can hide bad judgment if the test is too narrow. Include real examples.
Guardrail
A boundary that keeps AI behavior within the rules.
- What it does
- Blocks, flags, routes, or rewrites risky inputs and outputs.
- Why it matters
- AI systems need practical limits around privacy, brand voice, legal claims, safety, and approval.
- Use it for
- Subscriber emails, customer support, tool actions, private data, and any repeated task with real-world consequences.
- Tools to know
- Use private-context checklists, approval flows, policy reminders, redaction tools, and platform safety settings together.
- Watch out
- Guardrails are not magic. They need testing, logs, and a clear human fallback.
Fine-tuning
Training a model further on examples.
- What it does
- Adjusts model behavior using a curated set of task examples.
- Why it matters
- Useful when a repeated task has a stable pattern that prompting alone does not handle well.
- Use it for
- Consistent classification, extraction, formatting, or domain-specific response style at scale.
- Tools to know
- Most teams should start with examples, prompt templates, and evals. Fine-tuning is a builder path after that.
- Watch out
- Do not fine-tune before you have good examples, a clear eval, and a reason prompting is not enough.
