EasyHive

CLAUDE RUNS MY BIZ / 44

THE AGENT READINESS CHECKLIST

Is your AI agent real,
or a glowing orb?

The internet is full of people building their own Jarvis: holograms, glowing dashboards, a hundred agents supposedly running a whole company. It looks amazing. But looking busy and doing the work are two different things, and a name, a face or a job title in a prompt does not make a model an agent. This is the checklist we run before any agent gets near a customer: the eight questions, the prompt that audits an agent for you, and a safe first agent to set up in an hour.

Sources: Anthropic, Building effective agents and Effective context engineering for AI agents; OWASP Top 10 for LLM Applications 2025; Claude and Claude Code help docs, October 2026.

What actually makes it an agent

Anthropic's definition: an agent is a model that directs its own process and decides which tools to use. That only works when three things were set up on purpose.

THE PART	WHAT IT MEANS	WITHOUT IT
Memory somebody manages	It keeps what the job needs between runs (your rules, your customers, past decisions) in a place you can open, read and correct.	It forgets, repeats itself and contradicts you.
Its own tools	Only the apps and actions this one job needs. Nothing connected "just in case".	It cannot do the work, or it can do far more than you meant.
Real limits	Which files and accounts it can touch, where it may search, and which actions wait for your yes.	One bad instruction reaches everything you own.

WHY "JUST GIVE IT EVERYTHING" BACKFIRES Every request an agent makes carries its whole context with it: the instructions, the files you loaded and the conversation so far. Load everything and you pay for everything on every single request, so your budget or your plan's usage disappears fast. It also gets worse at the job: Anthropic's engineers write that as the context grows, the model's ability to recall what is in it goes down. Their advice is the smallest set of information that gets the job done, with the rest fetched only when it is needed.

The checklist: eight questions

Answer each in writing before it touches real customers, money or files. A blank answer is a no.

The job

What one outcome is it responsible for, and how will you check it did it well?

Good answer: "Reply to every new inquiry the same day with our prices and three times to book a call."

Red flag: "It helps with marketing." "It runs operations." Nobody can say what done looks like.

Memory

What must it remember between runs, where is that written down, and who reviews it?

Good answer: "Rules in one page of instructions, customer notes in one sheet. I review both every Friday."

Red flag: "It just remembers." Nobody can show you what it remembers or fix it when it is wrong.

Tools

Which apps and actions does this job need, and is every tool it has on that list?

Good answer: "Read the inbox, read the calendar, write drafts. That is all it is connected to."

Red flag: "We connected everything so it is flexible." Every extra tool is one more way to make a mistake.

Files and accounts

Which files, folders, projects and accounts can it open, and which are off limits?

Good answer: "Only the files in its own project. Finance, payroll and client contracts are blocked."

Red flag: It is logged in as you, with access to your whole drive, inbox and computer.

Where is it allowed to look things up: your FAQ, your website, the open web?

Good answer: "Prices and policies only from our price list and FAQ. The web only for public facts."

Red flag: It answers questions about your prices or policies from whatever it finds online.

What it needs to learn

What is the smallest set of information it needs for this job, and what can it fetch only when needed?

Good answer: "The FAQ, the price list and today's calendar. Old emails only when it needs one."

Red flag: "We uploaded all our files so it knows the business." Slower, pricier and less accurate.

Budget

Roughly what does one run cost, and where is the hard cap per month?

Good answer: "A monthly spend limit is set in the account, and I check usage every Monday."

Red flag: No cap anywhere. You find out what it cost when the bill arrives.

Your yes, the log and the off switch

Which actions wait for a person's yes, can you read what it did step by step, and can you stop it in one click?

Good answer: "Sending, refunds and deletes wait for me. Every action is logged. One toggle pauses it."

Red flag: "It runs on its own. We check when a customer complains."

HOW TO SCORE IT All eight answered: run it for one week on real work while you watch. Any blank on 04, 07 or 08: stop there. Those are the three that reach your customers, your money and your files.

The audit prompt

Paste it into a new Claude chat with whatever describes the agent: its instructions, its tools, screenshots of its settings, or the sales page of a tool you might buy. Fill in every [BRACKET].

THE AGENT AUDITCOPY EVERYTHING IN THIS BOX

You are auditing an AI agent before a small business trusts it with real customers, real money and real files. Be blunt and specific. Use only what I give you below. When something is not stated, mark it UNCLEAR and never assume it is fine. My business: [WHAT YOU SELL AND TO WHOM] The one job I want this agent to do: [THE OUTCOME, e.g. reply to new inquiries and book calls] How it is set up or sold: [PASTE THE INSTRUCTIONS, THE CONNECTED TOOLS AND THEIR PERMISSIONS, OR THE SALES PAGE] Check these eight areas. For each one give YES, NO or UNCLEAR, one line of evidence quoted from what I gave you, and the risk in plain words: 1. Job: one outcome, and a way to check it was done well 2. Memory: what it keeps between runs, where it is written, who reviews it 3. Tools: only the tools this job needs, nothing extra 4. Files and accounts: what it can open, and what is off limits 5. Search: where it is allowed to look things up 6. Context: the smallest set of information it needs, the rest fetched only when needed 7. Budget: rough cost per run, and a hard spending cap 8. Human control: which actions wait for a person's yes, a step by step log, and a one click off switch Then give me: A. A one line verdict: a working agent, a promising demo, or a glowing orb. B. The three changes that would make it safe to run on real customers, most important first. C. Which of its limits live only in the instructions (a request) and which are enforced by a setting or a permission (a lock). D. A one week test: five real tasks to give it, and what I should check after each one. E. The questions to ask the seller or builder before I pay for it or switch it on. Write for a business owner, not a developer. Short sentences, no jargon.

A LIMIT IN THE PROMPT IS A WISH Writing "never email a customer" in the instructions is a request, not a lock. Claude Code's own docs say it plainly: instructions shape what Claude tries to do, permission settings decide what it is allowed to do. Every hard limit belongs in a setting.

HOW TO READ WHAT COMES BACK Start with part C and move every limit that is only a request into a setting (page 4 shows where). Then make the first change in part B, and only that one. Run the audit again after you fix it: a second pass with fewer UNCLEAR answers is your proof that it worked.

BUYING AN AGENT INSTEAD OF BUILDING ONE? Paste the sales page and send part E to the seller before you pay. If they cannot tell you which permissions it asks for, where its memory is stored, how to cap the spend and how to see what it did, you are buying the glowing orb.

Where each answer lives in Claude

Turn your answers into settings. Left column if you work in the Claude app, right column if you build in Claude Code.

QUESTION	IN THE CLAUDE APP	IN CLAUDE CODE
Job and rules	Create a Project for the agent, then Set project instructions with the job and the rules.	A `CLAUDE.md` in the project folder, or the agent's own file in `.claude/agents/`
Memory	Each project keeps its own memory, separate from your other chats. Review and edit what it remembers in Settings > Memory.	`CLAUDE.md` holds what you write, auto memory what Claude learns. Open both with `/memory`
Tools	Customize > Connectors, pick the connector, then set each tool to Always allow, Needs approval or Blocked.	The `tools:` line in the agent file, plus allow, ask and deny rules in `/permissions`
Files	Inside a project it works from the knowledge you add there, plus only the connectors you switch on in the + menu.	Deny rules such as `Read(./finance/**)`. Deny always wins over allow
Budget	Your plan's usage limits. If you turn on usage credits, set a monthly spend limit in Settings > Usage.	A monthly cap on the workspace's Spend limits tab in the Claude Console; `--max-budget-usd` for scripted runs
Off switch	Turn the connector off in the + menu, or disconnect it in Customize > Connectors.	`Esc` stops it mid task. A deny rule with a bare tool name removes that tool entirely

What to do in the next hour

Build the safest useful first agent there is: one that reads your inbox and drafts replies, while you press send.

TIME	DO THIS	WHY
10 min	Answer the eight questions for one job: replying to new inquiries.	The answers become its instructions
10 min	Create a Project called Inbox Helper. Paste your answers into the instructions and add your FAQ and price list.	Its own memory and only the files it needs
5 min	Connect Gmail. Set search and read to Always allow, sending, replying and forwarding to Needs approval.	It can read and draft, never send on its own
20 min	Ask it to draft replies to your last five inquiries. Compare each one with what you actually sent.	A real test on real work
15 min	Run the audit prompt on your own setup and make the first change it suggests.	Your agent passes its own checklist

WANT MORE LIKE THIS? Every week we take one piece of how we run our business with Claude and turn it into something you can use the same day. The rest of the free resources are in the Easy Hive community classroom.

Is your AI agent real,or a glowing orb?

Is your AI agent real,
or a glowing orb?