Our own directory files them under those three names: agents, skills and MCP servers. Which one is the "real" agentic AI matters less than which one you could turn on this week, and what would change if you did.
In one line each
- Agent
An AI system that runs a job end to end: it plans the steps, uses tools to do them, checks the results, and comes back when the job is done or it needs a decision from a person.
True on October 1, 2026. Source: Anthropic, Building effective agents
- Skill
A written playbook, a folder with a SKILL.md file in it, that an AI assistant loads when your request matches what the skill says it's for.
True on October 1, 2026. Source: Anthropic, Agent Skills overview
- MCP server
A connector built on the Model Context Protocol, an open standard, that lets an AI assistant look things up and take actions in another piece of software.
True on October 1, 2026. Source: Model Context Protocol docs
An agent runs the whole job#
A chatbot
- YouYou ask
- The AIIt answers
- YouYou read it and ask the next thing
- The AIIt answers
- YouYou ask again
- The AIIt answers
3 of 6 steps are yoursYou type every next step.
An agent
- YouYou hand over the job
- The AIIt plans the steps
- The AIIt uses a tool
- The AIIt checks the result
- The AIIt does the next step
- YouYou hear back: done, or a decision only you can make
2 of 6 steps are yoursYou hear back at the end, or at a decision only you can make.
A chatbot answers, then waits for your next message. An agent picks the next step and takes it itself: it plans a sequence of steps, uses tools to carry them out, checks its own output, and comes back when the job is done or it's stuck. Anthropic, which makes Claude, draws the line the same way: agents "dynamically direct their own processes and tool usage" (Building effective agents, 2024-12-19, checked 2026-10-01).
The difference shows up in what you do while it runs. With a chatbot, you're the one typing the next instruction. With an agent, you're the one who gets pinged when it's finished, or when it hits a decision only a person can make.
45 tools filed under AI Agents in our directory today, including:
The AI Agents categoryA skill is a playbook, not an app#
---name: invoice-remindersdescription: Drafts reminders for overdue invoices. Use when someone asks to chase or follow up on unpaid invoices.---# Invoice reminders1. Pull the invoices more than 14 days overdue.2. Draft one reminder each, in our usual tone.3. Never send. Leave every draft for a person.Chase this week's overdue invoices
Loaded skill: invoice-remindersFound 3 invoices more than 14 days overdue. Three drafts are below for you to read. Nothing has been sent.
Draft 1Draft 2Draft 3A skill is a written set of instructions an assistant reads before doing a job: a folder with a file called SKILL.md at the top. You add it once. After that you don't switch tools to use it. You ask for the job in plain words, the assistant matches your request against the skill's description, and it loads the steps inside the conversation you're already having (Anthropic's Agent Skills docs, checked 2026-10-01).
That also means a skill is only as good as what's written down in it. Anthropic's own advice is to use skills "only from trusted sources: those you created yourself or obtained from Anthropic", and to audit every file in any other skill before you use it (same docs).
An MCP server is a wire to what you already run#
You askin plain words
Your assistantClaude, ChatGPT
Notion MCPMCP server
Stripe MCPMCP server
Wix MCPMCP server
- no copy and paste
Notionlooks up or acts
Stripelooks up or acts
Wixlooks up or acts
The first two are about what the AI does. This one is about what it can reach. An MCP server connects an AI assistant to software you already run, a calendar, a CRM, a spreadsheet, so the assistant can look something up there or take an action there directly, instead of you copying numbers back and forth. MCP is an open standard, and its own docs compare it to a USB-C port for AI applications (modelcontextprotocol.io, checked 2026-10-01).
30 MCP servers in our directory today, including:
Every MCP serverThe failure worth guarding against#
Seven AI models, $300 each, a planned 72 hours to run a business
$12,431
in invoices sent to strangers for unsolicited work, all later voided
2,797
spam emails sent
$0
in revenue
None of this is set it and forget it. In a test Bottleneck Labs published, seven leading AI models each got $300 in a real bank account and a planned 72 hours to run a business on their own. The lab logged every message and tool call, halted two models early and gave one 12 more hours. One model sent 50 invoices, from $49 to $599, to strangers for unsolicited work. The lab found out when people emailed it about the spam, then halted that run and voided every invoice. Between them the models made no money. Their bank balances went from $2,100 to $1,740.20, and the AI usage to run them cost about $2,800 more (Bottleneck Labs, checked 2026-10-01; discussed on Hacker News, 101 points, 2026-09-07). That's one lab's test, and we haven't rerun it. The shape of the risk needs no rerun: the invoices and emails reached real people before anyone stopped them.
The rule we give clients is below. It's our recommendation, not a standard.
Fine to run on its own once set up
Looks, drafts, sorts
- Look something up and report back
- Draft a reply for you to read
- Sort, summarise and file
Waits for you until you've watched it work
Money and customers
- Send money or an invoice
- Send anything to a customer
It's the rule that already applies to a new hire: nothing that sends money or talks to a customer runs unsupervised until you've watched it do the job with you standing there. Anthropic's guide sets no rule like this, but it does say agents "can then pause for human feedback at checkpoints or when encountering blockers" (Building effective agents, checked 2026-10-01), so a checkpoint before money or customers is something an agent can be built to stop at.
Try one this week#
Pick a job you already do by hand every week, and find its card. If you could describe it start to finish, look for an agent that does it. If it's something you already know how to explain to a new hire, write it up as a skill's instructions and see if your assistant loads it. If it's mostly "go check X and tell me," add an MCP server for X instead of copying and pasting between the two.
Before you build any of these, work out what it's worth. Our free automation ROI calculator does the math from your own hours and costs, payback time included.