«I’ve built a skill for cashing up.» «It’s hooked up to Drive through MCP.» «I left it on a /goal and it was done by morning.» If you read about AI agents, lines like these turn up in every paragraph. Here is what MCP is in simple terms, what a harness is, what an AGENTS.md file is for and the rest of the words around agents, each in a couple of sentences with an office example.

One trick to make it all fit: picture the agent as a very bright new hire who knows nothing about your company. Almost all of these words are ways of explaining how you work or of setting limits.

What an agent, a model and a harness are in AI

Model. The engine: Claude Opus 5.5, GPT-6 Sol, Gemini. On its own it reads and writes, but it can’t open your files or press a single button.

Agent. A model inside a loop, with tools, that looks, decides the next step, does it and checks whether it worked, until it’s finished. You tell it «put these 80 delivery notes in a spreadsheet and check the totals add up» and come back a little later to review; I go through it slowly in what is an AI agent.

Harness. The program that wraps the model and turns it into an agent, giving it tools, memory, rules and the working loop. Claude Code and Codex are harnesses: as Anthropic itself puts it, Claude Code is the harness and Claude is the model inside it. That’s why the same model performs differently depending on the program.

What a token is, and what the context is

Token. The piece of text a model works with, sometimes a word and sometimes part of one. On a subscription you don’t see them, but they count: according to Anthropic, what you use also depends on the size of the files you attach, so summarising a 200-page PDF eats far more of your limit than a question. Through the API you pay for them: Claude Opus 5.5 charges $4 per million you send it and $20 per million it writes, plus VAT.

Context. Everything the model has in front of it at once: the conversation, the files it has opened and its instructions. It’s its desk, and although Claude Opus 5.5 takes a million tokens (about 555,000 words, according to Anthropic), a long agent session fills it.

Compaction. When the desk fills up, the agent summarises the older part to carry on, and details get lost in that summary. If you only said «amounts always without VAT» in the chat, it may forget; whatever has to last, put it in writing in AGENTS.md.

What AGENTS.md is, and AGENTS.md vs CLAUDE.md

AGENTS.md. A text file you leave in your work folder with what you’d tell someone new on their first day: how you work and what’s off limits. The agent reads it at the start of every session. OpenAI launched it in August 2025 with other companies, Google and Cursor among them; today the Agentic AI Foundation, part of the Linux Foundation, looks after it, and Codex, Cursor, GitHub Copilot and many more read it.

CLAUDE.md. Claude Code’s equivalent, which since 18 September 2026 reads the AGENTS.md when the folder has no CLAUDE.md (if both are there, it only reads CLAUDE.md). If you use Claude Code and Codex, a single AGENTS.md works for both.

The easiest way is to let the agent write it. Open it in your test folder and ask for it with your own details:

Create an AGENTS.md file in this folder with these instructions, and ask me if you think something important is missing.

- We are a carpentry workshop in Zaragoza that makes furniture to measure. There are three of us.
- Write to clients in Spanish from Spain, in a friendly, professional tone. Amounts like 1.234,56 €.
- Quotes go in quotes/, named client-date.pdf.
- Never delete files: anything we don't need goes to the bin/ folder.
- Don't open the payroll/ folder.
- Before you call a quote finished, check that the total is the sum of the line items plus VAT.

And when you’ve corrected the same thing twice, ask it to add it there.

What Claude skills (and everyone else’s) are

Skill. A folder with a SKILL.md file that explains step by step how a particular task is done, like the procedure you’d hand someone so they cash up exactly the way you do. At the start the agent only reads each one’s name and description, and opens the full skill when you ask for something that fits, so it can keep lots of them without cluttering its desk. Anthropic published it as an open standard in December 2025, and today it works in more than forty tools, including Codex, Cursor and GitHub Copilot.

You don’t need to write it yourself either. Do the task once with the agent, fix whatever goes wrong and ask:

Turn what we've just done into a skill called monthly-report, so next time you do it the same way. Before you create it, show me the steps it will have.

Plugin. Several skills, connections and rules packaged so you can install them in one go in Claude Code, Codex or ChatGPT. Only from sources you know, because they put instructions into your agent.

What MCP is in simple terms, and what a connector is

MCP. It stands for Model Context Protocol, the standard for plugging an agent into other programs, such as your email, your calendar, Google Drive or your invoicing software; Anthropic itself compares it to a USB-C port. Anthropic created it in November 2024 and in December 2025 donated it to the Agentic AI Foundation, which also hosts AGENTS.md; today ChatGPT, Claude, Gemini, Copilot and many agents use it.

Connector. A connection to another program, such as your calendar. Some use MCP and others do not. Before asking it to find two free slots this week, check which account you connect and which permissions you grant.

So how does a skill differ from MCP? With MCP the agent can get into a program. A skill explains how a task is done, and it often draws on those connections: the Drive connector opens your sales spreadsheet and the monthly report skill tells it what to do with it. Only connect what you use, one at a time.

Permissions, Manual mode and checkpoints: how you set limits

Permission modes. They determine what can run without your approval. In Claude Code, Manual retains pre-approved actions; Auto adds automated review, which can fail. In Codex, check the workspace scope and approval for additional access. The mode name alone is not enough: review its rules and do not expand access just for convenience.

Manual mode. The one I’d use for the first month, always with a test folder and copies. If it asks to run something like python read_invoices.py and you don’t know what that is, reply «explain what that does first»; how to set it without touching the terminal is in Claude Code and Codex for non-coders.

Checkpoint. A copy Claude Code saves every time you send it a message, so you can go back with /rewind or by pressing Escape twice. It only undoes what it edits with its own editor, not what it moves or deletes with commands, so you still need your backups.

Hook. A check triggered by a specific event, outside the conversation memory. It can block an action when configured to do so. Test which events it covers and what happens on failure before relying on it to protect files.

Prompt injection. Orders hidden in something the agent reads, such as a web page or a PDF, to make it do something you didn’t ask for. It’s the risk most specific to agents, and the reason deleting, sending, paying or publishing are always your call.

Subagents, /goal and routines: when it works alone for longer

Subagent. A helper the main agent sends off to do part of the job: one goes through the 40 rental contracts, another through the receipts, and the main agent puts together what they bring back. They run in parallel, and each one uses up your limit.

/goal. Sets a verifiable objective in Claude Code or Codex. For example: identify invoices, payments and differences without inventing adjustments, and provide evidence. It does not guarantee completion or change permissions. Limits, errors or blockers can stop it; review its state and do not confuse an agent verdict with an independent check.

Routine or scheduled task. A job that repeats on its own, like «every Monday at 8, prepare the sales summary». Claude Code routines run in the cloud and don’t need your computer, but they can’t see your files: only what’s on GitHub, your connectors and the websites you allow. If the work lives in your files, use the scheduled tasks in the Claude or ChatGPT desktop app, with the computer on and the app open.

Usage limits and resets: what you spend

Usage limit. What your plan lets you use before you have to wait: Claude and Codex both have one that renews every five hours and a weekly one, and in Claude you see them under Settings, Usage. When you hit the ceiling, the agent stops until it renews, unless you buy usage credits.

Reset. A button that refills your limit in one go; Anthropic and OpenAI give them away now and then, the ones you can save expire, and in Claude they appear under Settings, Usage. I’d keep one for when the limit catches you halfway through something urgent.

Which ones you need in your first month

You don’t need to learn them all. I’d go like this:

  1. Week one: agent, Manual mode and a test folder. That’s enough to work.
  2. Once you’ve corrected the same thing a couple of times: an AGENTS.md.
  3. When you repeat a task every week: a skill.
  4. When you want it to use your email or your Drive: a connector, one at a time.
  5. Later on: /goal and routines.

You can park A2A without guilt (a standard for agents from different companies to talk to each other, created by Google and now at the Linux Foundation), along with worktrees and headless mode. And if you want a day-by-day plan to get this far, it’s in learn AI without coding.

Sources, checked on 26 September 2026: MCP announcement, MCP donated to the Agentic AI Foundation, Linux Foundation press release, AGENTS.md, Claude Code changelog, Claude Code glossary, Claude Code permission modes, Codex permission modes, Agent Skills, Claude API pricing, Claude usage limits and what a limit reset is.