Best AI Agent for Small Business: One Folder, Three Tools

Every list of "the best AI agent for small business" answers a question you did not ask.

You asked: which one should I use. What you actually meant was closer to this: what if they raise the price, shut it down, or ban my account? Those are different questions. A feature comparison answers the first one and quietly ignores the second.

Here is why the second one deserves an answer. Cursor, a tool a lot of people had been feeding notes into for months, removed its entire Memories feature in a minor version update. Not wound down over a year — removed. The advice from staff on the company's own forum was to export what you had and paste it into plain text rules files by hand. I checked the documentation page for that feature again on 31 July 2026: cursor.com/docs/memories returns a 404.

Nobody did anything wrong. A product team looked at a feature, decided it was not worth keeping, and shipped. Normal product decision — and also the exact shape of the thing you are afraid of, on a paid tool with a lot of users.

So the real question is not which one is smartest this quarter. It is: if the tool you use goes away next spring, how much of your setup comes with you, and how much do you rebuild from nothing? This article is the check for that. One folder of your own rules, three different places to run it, and four things to write down each time.

One thing to be straight about before you read on. This piece has the method, the files, and everything I could verify from vendor documentation on 31 July 2026. It does not have my own run logs yet, because I have not done the three runs. Every place a real number belongs is marked. An honest gap beats a plausible figure I made up.

The question the lists don't answer

I searched the phrase this article is named after and read what came back. Ten results, one shape: a numbered list of products, each with what it is best for, what it costs, sometimes a security badge. Three of the top results were written by companies that put their own product at number one.

That format is not dishonest. It is built on an assumption that stops being true about a week after you start — that you are choosing from zero. You are not. If you have followed this series you now have instructions you wrote, job files with your actual steps in them, a checklist for finishing work, and a page listing the jobs you will not hand over. That took real hours, and none of the lists tell you what happens to those hours when you move.

There is a second gap, and it is bigger. Those lists are all about capability, and capability is the thing that changes fastest and matters least to you. The model that drafts the best engagement letter today will not be best in March, and you will not notice, because the difference between "good" and "slightly better" on a two-page letter is not something you can feel. What you can feel is spending a Saturday re-teaching a new tool the fourteen things about your practice that the old one already knew.

That cost is the one worth measuring, and moving is the only way to measure it.

A different test: can you leave?

The check is simple enough to describe in one line. Take the same folder to a second place and do the same real job.

Not a demo job. A job you did last week and can grade, because you know what the right answer looked like. Pick one of these, whichever is closest to your work:

• Front of house. A new client comes in. Before anyone hears the story, the conflict check runs — existing clients, former clients, and people who consulted you and never signed. The American Bar Association's own guidance on intake puts it in that order for a reason, and Model Rule 1.18 is why the third group is on the list. If you are a real estate agent, the equivalent is the written buyer agreement that has had to be signed before a showing since 17 August 2024 under the NAR settlement terms.
• Delivery. The walk-through before you call it done. The contractor prepares the list of items to complete or correct, the architect verifies and amends it, and then the owner, architect and contractor all sign — that is the sequence in AIA A201-2017 and the G704 certificate. If you translate, the equivalent is the revision step in ISO 17100, which has to be done by a second person; you are not permitted to revise your own translation.
• Books. Month end. Ledger balance, the sum of every individual client ledger, and the adjusted bank balance all have to agree. In North Carolina, Rule 1.15-3 requires the lawyer to sign and date the reconciliation and keep it six years, and requires monthly review of bank statements and check images — a duty the state bar's handbook says explicitly cannot be delegated.

Pick one. Run it in the tool you use now. Then run it in a second tool with the same folder attached and nothing else changed. Write down four things. Done.

Notice what all three jobs have in common: none of them is hard, and all of them are ordered. That ordering is where the money is. In the American Bar Association's study of legal malpractice claims from 2016 to 2019, administrative errors accounted for 19.59% of claims — 7.4% from calendaring alone, 4.08% from clerical mistakes — and client relation errors for another 16.7%. Conflicts of interest were 4.9%. Add it up and roughly four claims in ten had nothing to do with knowing the law. They were sequence and follow-through.

A written-down process is good at exactly that kind of work — which is why moving one without damage is worth checking properly.

This check works better than a feature comparison because it produces a number you can act on. "Two hours to move, output was the same except it dropped the deadline column" is a decision you can make. "Best for teams that need advanced automation" is not.

What you're actually moving

Here is the folder. Nothing in it belongs to any vendor — plain text, written in your own words.

my-work/
├── AGENTS.md            # read this first
├── jobs/
│   ├── new-client-intake.md
│   ├── delivery-check.md
│   └── month-end-books.md
├── checks/
│   ├── before-sending.md
│   └── switching.md     # new in this piece
├── boundaries.md        # jobs that do not get handed over
└── memory/
    └── decisions.md     # things we settled, and when

AGENTS.md is the file most agent tools look for when they start. Nobody invented it for this article. OpenAI published it in August 2025 and donated it to the Agentic AI Foundation under the Linux Foundation on 9 December 2025 — the same foundation Anthropic handed its own connector standard to, with OpenAI, Anthropic and Block as founding contributors. There is no schema, no required fields and no version number. The project's own FAQ says so in one line: "No. AGENTS.md is just standard Markdown. Use any headings you like; the agent simply parses the text you provide."

On 31 July 2026 the project site listed 23 named tools that read it, and stated that it is used by more than 60,000 open-source projects — a figure published at the December 2025 donation announcement that the site has not revised since. Treat it as a floor, not a current count.

The one file I want you to add before you try any of this is checks/switching.md. It turns the check into something you keep rather than something you remember. Copy it as is:

# Switching check

Run this whenever we move this folder to a different tool, or when
the tool we use changes its plan, its price, or its limits.

## The job we test with
Always the same one: <name the job file here>.
Always the same input: <name a real past matter/project/month>.
We know what the right answer looks like, because we did it by hand.

## Before running anything, ask the new tool four questions
1. Which of these files can you see right now? List them by name.
2. Do you read all of them on every message, or look things up as needed?
   If you do not know, say you do not know.
3. Read AGENTS.md and everything in checks/. Name every instruction in
   them you cannot follow here, and say why.
4. What do I need to paste or re-upload because it did not come across?

## Then run the job and record four things
- Time to connect: from first upload to first usable answer.
- What changed in the output, compared with the known-good version.
- What it cost: plan, per-message charges, anything metered.
- Where it fell short: the specific step it got wrong or skipped.

## Rules
- If it got everything right, we did not test hard enough. Pick a
  harder input and run it again.
- Anything in boundaries.md still gets done by a person. Moving tools
  does not move responsibility.
- Date and sign this page. Keep the old versions.

Last run: <date> by <name> on <tool>.

Then add this section to your AGENTS.md. Six lines, and it is the difference between a folder that explains itself on arrival and one that needs you standing next to it:

## If you are reading this for the first time

You are picking up work that was set up somewhere else. Before you
start: read boundaries.md, then the job file I name. If anything in
those files describes a tool feature you do not have, tell me which
line, and do the rest. Do not substitute your own steps for the ones
written down.

That is it. Two short files. The rest of the folder you already wrote in the earlier pieces of this series, including splitting one long set of instructions into separate job files and the list of jobs you must not hand over.

Things you can know before you move anything

Some of the answer does not require running anything — the vendors have written it down, and it explains most of the friction people hit. I checked all of this on 31 July 2026.

Three things are worth pulling out of that table, because they change how you plan the move.

Neither of the two big web assistants promises to read all your files every time. Both describe some form of looking things up. Sensible engineering, genuine problem for you: the rule you care most about might sit in the file it decided not to consult. Hence the four questions in switching.md starting with "which files can you see" — you want that on the record before you judge the output.

The nesting behaviour is not standard even though the file is. The project FAQ says the closest file to the work wins. In practice Codex concatenates every file from the root down and relies on later text overriding earlier text; Cursor merges with the more specific one winning; Zed takes the first match and stops. Same file, three different meanings of "wins". If your folder has one AGENTS.md at the top you will never notice. The moment you add a second one in a subfolder, you will.

The way these tools read each other's files is hand-built, not agreed. opencode, Zed, Warp, the server-side version of GitHub Copilot, Devin's command line tool and Amp will all read a CLAUDE.md if they find one. Claude Code does not read AGENTS.md. Tools support each other by teaching themselves each other's filenames, one at a time. It works. It works by goodwill rather than by rule, and that is worth knowing before you rely on it.

Three runs, one folder

This is the part where I owe you numbers and do not have them yet.

Run one — the web assistant you're on now

Create a project. Upload the folder contents one file at a time, because you cannot upload the folder. Paste the four questions. Then run the job.

What is worth watching: whether it lists every file back to you, and whether it names the boundaries file without being told. If the answer to question 2 is some version of "I look things up when relevant", write that down — it is the setting your output was produced under.

[NEEDS REAL RUN: time from first upload to first usable answer on a web project, plus the tool's verbatim answers to the four questions — requires uploading the real folder and logging the session]

Run two — the other web assistant

Same folder, different product. Here the file count cap matters. If your folder has more files than the plan allows, you have to decide what to leave behind before you have any evidence about what matters — which is itself a finding worth writing down.

[NEEDS REAL RUN: time to connect on the second web project, and whether the file count cap forced anything to be dropped — requires a real account at a stated plan level]

Run three — a command line tool

Different in kind: you upload nothing. The folder sits on disk and the tool reads AGENTS.md where it finds it. If you have never used one, this run will feel strangest and finish fastest, because there is no upload step at all. I am not expecting you to move to a terminal — it is in the test because it shows whether your folder is readable by something that was never designed around it, which is the whole claim being checked.

[NEEDS REAL RUN: setup and first-answer time on a command line tool, and whether the 32 KiB instruction cap was hit silently — requires running the real folder and inspecting what actually loaded]

The four columns

Here is the sheet. Print it, or paste it at the bottom of checks/switching.md.

[NEEDS REAL RUN: my own filled-in version of this table — three runs of the same job on the same input, with timings from first upload to first usable answer, a step-by-step comparison against the known-good version, the actual charges, and the specific step each one got wrong]

One instruction about that last column, and it is the one people skip. If all three come back clean, the test failed. Three different products, three different ways of deciding which of your files to look at — if none of that shows up in the output, you tested something too easy. Go back and pick the input you got wrong last time: the one with the awkward second party, or the month with the correction in it.

A comparison where everything passes is a comparison nobody learned anything from. The gap is the finding.

What comes across, and what stays behind

Here is the part I can answer with confidence, because the vendors have written it all down.

What comes across: everything you wrote. AGENTS.md, the job files, the checklist, the boundaries page. Text goes anywhere text goes. The 23 tools on that list read the file natively; two more read it with a line of configuration; the two big web assistants read it as an uploaded document. Nothing about that depends on any company staying in business.

What stays behind: everything the tool worked out about you on its own. Every vendor has some version of this now, and none of it travels. The evidence that it will not change soon is unusually clear:

• ChatGPT says its memory summary does not contain everything it has remembered from your chats; to find out whether it knows something you have to ask it in conversation. You cannot copy what you cannot see.
• Copilot's user-level preferences are owned by the billing entity that issued your licence, and are only retrieved under the currently active one. Change employer and they are gone, same person, same machine. Facts and preferences that go 28 days unused are deleted automatically.
• Claude's exported account data cannot be imported into another Claude personal account. Not into another vendor — into another account at the same vendor.
• The import features that Claude and Gemini both ship for pulling memory across from a rival work by handing you a prompt, asking you to run it inside the competitor's product, and having you paste back whatever prose it produces. Claude labels its version experimental and says it may not successfully absorb what you import.

Two large companies looked at the problem of moving memory between products and shipped copy-and-paste as the answer. No criticism of either. It tells you there is nothing else available to build on. The only specification I could find that proposes a common format was published in February 2026 by a single author, with no standards body and no vendor behind it — and it says plainly that vendors do not support it.

So the useful conclusion is the one OpenAI's own Codex documentation reaches, which is worth reading twice given who wrote it: put rules that must always apply into AGENTS.md or a document you keep and version yourself, and treat the tool's memory as a helpful recall layer rather than the only home for anything that matters.

That is the argument for writing things down, made by a vendor about its own product. The full breakdown of what each one's memory actually holds is in the piece on what a vendor's memory actually holds.

One more thing stays behind, and it surprises people: the project itself. Your files come back because you have them on disk. The container does not — not the instructions attached to it, not its own memory space, not the way the conversations were organised. Google's own import documentation says it cannot bring project files across at all.

Switching is not free

This is not a two-minute job, and selling it that way would cost you a Saturday and cost me your trust in everything else here. Here is what actually needs re-doing, based on what is documented rather than what I wish were true:

The file has to arrive in the right shape. Capitals matter to at least one tool. A second file in the same folder can beat yours in at least two. If your folder has more files than the new plan allows, you are cutting before you have evidence.

Anything that leaned on the old tool's memory has to be found and written down. This is the expensive part and it is invisible until you switch. The old tool knew that you always copy the assistant in, that one long-running matter is billed differently from the rest, that nothing goes out on a Friday afternoon. You did not write those down because you did not have to. The new tool does not know any of them, and you will not discover the list by reading — you will discover it one wrong output at a time. Every one you find, add to memory/decisions.md, and you never pay for it again.

Size limits differ enough to change behaviour. A folder that fits comfortably in one place can be silently truncated in another. Codex stops at 32 KiB by default without saying anything. Windsurf caps its global rules at 6,000 characters and workspace rule files at 12,000. If your instructions are long, the second tool may be working from part of them.

The output will not be identical, and some of the difference is real. Different models write differently. Some of that you will not care about. Some of it — a dropped step, a softened caveat, a date format that breaks a downstream form — you will care about a lot. Hence the third column.

[NEEDS REAL RUN: how long the re-tuning actually took, broken down by what caused it — requires doing the move and keeping a log rather than estimating afterwards]

Plainly: the part that moves is the part that took the most thinking, and the part you rebuild is the part you never consciously built. Better trade than the alternative, and still not free.

When it's worth it, and when it isn't

Most of the time, don't. Switching for its own sake is a hobby, and you have a business.

Worth doing:

• The price moved and the sums changed. Not "it went up" — run your actual monthly volume against the new pricing and see if it still makes sense.
• You are hitting limits during work you cannot pause. Rate caps that bite in the middle of a filing week are a different problem from rate caps that bite on a Sunday.
• One specific job is clearly better somewhere else. Not everything — one job. It is completely reasonable to keep the books work in one place and move the drafting to another. The folder does not care; the job files are separate for exactly this reason.
• Something in the product's direction stopped matching yours. A feature you built around got removed. A data policy changed. Legitimate reason, and you do not owe anyone a longer explanation.

Not worth doing:

• A new model scored higher on a benchmark. Benchmarks are not your intake process.
• Someone online says the other one is better. They are not doing your job with your files.
• You are bored. Fair — but do it on a quiet week, not before month end.

The best AI agent for small business, in the only sense of "best" that survives a year, is the one where the answer to "can I leave" is yes and you have checked recently. Not the one with the longest feature list. The one you have already proved you can walk away from.

Which is also, quietly, the answer to whether it is safe to pay for anything set up this way. If the setup lives in files you own, buying a head start does not lock you in — you are buying the files, and the files come with you.

The receipt

Everything I checked, how I checked it, and what I got wrong on the way.

Checked on 31 July 2026, directly from the source page:

• The agents.md site listed 23 named tools. Claude Code was not among them. The stated adoption figure was still "over 60k open-source projects", unchanged since the December 2025 announcement.
• The two FAQ answers quoted above are verbatim from that page.
• Claude Code's documentation states "Claude Code reads CLAUDE.md, not AGENTS.md" and gives two ways round it: an import line, or a stand-in file pointing at the real one.
• Claude Code's auto memory documentation states files are not shared across machines or cloud environments. Even when the memory is a file, syncing is your problem.
• cursor.com/docs/memories returned HTTP 404.

Wrong answers I rejected while writing this:

1. "AGENTS.md has a schema you should validate against." It does not. There is no template to conform to, nothing that checks the file, no required fields, no version number. At least one widely-circulated guide describes a canonical JSON schema, lifecycle events and a registry. None of that exists, and it contradicts the project's own FAQ. If a guide tells you to validate the file, close the guide.
2. "Claude Code reads AGENTS.md as a fallback." Two well-ranked articles say this. The official documentation says the opposite in one sentence. When a secondary source and a primary source disagree about the primary source's own product, believe the product.
3. "Nobody uses this outside software." I wanted this to be true because it makes the angle cleaner. It isn't. There is a public bookkeeping project on GitHub with an AGENTS.md at the top of it, and a fair number of people using one to run their personal note collections. What is true, and is enough: the format defines itself as being for coding agents, every tool that reads it is a code editor or a terminal, and no professional body, standards group or firm has published a method for using it as an operating manual. That is a gap, not a void. The distinction matters and I nearly wrote the wrong one.
4. Recomputing the adoption number myself. I tried querying GitHub's code search API with the same query the site links to. It returned a number so far off that it is clearly measuring something else. The count on the site is a file count, not a project count — one large project can contribute dozens on its own. I left the published figure with its date attached rather than substituting a number I could not stand behind.

What I have not tested, and you should hold me to:

• I have not run the three-way comparison. No timings, no bills, no before-and-after comparisons. Every one of those places is marked in this article and will stay marked until the runs happen.
• I have not tested what happens when the folder has more files than a plan allows.
• I have not tested the nesting behaviour with a second AGENTS.md in a subfolder across all three, which is where I expect the differences to be sharpest.
• I have not had any of the job files graded by someone who does that work for a living. Until that happens, treat the file structure as sound and the domain content as a starting point.

I'm not a lawyer, a contractor, or a bookkeeper

I don't do any of the three jobs in this article, and I am not going to pretend otherwise.

What I bring is the opposite of expertise, and I think it is more useful here than it sounds. Someone who has run intake for fifteen years asks how AI can make step four faster. I ask why step four exists, because I do not know, and half the time the answer turns out to be that it exists to catch something a written check catches better.

That has a hard edge though. I will tell you how to get a process to run somewhere else. I will not tell you whether the output meets your obligations. The lawyer signs the reconciliation. The contractor swears the affidavit. The translator certifies the translation. None of that moves because your files moved.

So: I have written the switching check, and the file structure it runs against. I have not yet run it on a real matter. If you do this work and there is a step I have obviously misunderstood, tell me — that correction is worth more to me than the article is.

Copy this: the switch test

Paste this into the new tool after you have attached the folder and before you ask it to do anything.

I am moving my work folder from one AI tool to another. The folder is attached
(or in this project). Do not start the job yet.

First, tell me four things, in this order:

1. WHICH FILES YOU CAN SEE. List every file you actually have access to right
   now, by name. If you cannot see a file I mentioned, say so plainly.
2. WHICH FILES YOU WILL READ EVERY TIME. Tell me whether you load all of these
   on every message, or look things up only when they seem relevant. If you do
   not know, say you do not know. Do not guess.
3. WHAT YOU CANNOT DO HERE. Read AGENTS.md and the checks folder. Name every
   instruction in them that you are unable to follow in this tool, and say why.
4. WHAT YOU WOULD NEED FROM ME. List anything you need me to paste, re-upload,
   or re-explain because it did not come across with the files.

Rules: do not invent capabilities you do not have. Do not tell me a step
"should work" — tell me whether it does. If a rule in boundaries.md says a human
must do something, repeat that rule back to me instead of offering to do it.

After you answer all four, stop and wait.

Then, and only then, run the job:

Now run the job in jobs/<the job file I name>. Follow it exactly as written.

When you finish, add a short section at the bottom called "Where I had to guess",
and list every point where the job file did not tell you enough and you filled in
the gap yourself. Quote the step number. If there were none, write "none".

Do not improve the job file. Do not skip steps you think are unnecessary.

That second section is the one to read first. Every guess it admits to is a line missing from your job file, and fixing it makes the file better in every tool, not just this one.

Your next thirty minutes: create checks/switching.md from the text above, add the four-line section to your AGENTS.md, and pick the one real job you will always test with. That is the whole setup. The run itself can wait for a quiet afternoon.

Questions people ask

What is the best AI agent for a small business?
There is no answer that stays true for more than a few months, which is why every list you read is out of date by the time you act on it. The more useful question is which one you can leave. Write your rules into a folder of plain text files, then confirm a second tool can pick up that same folder and do the same job. Any tool that passes is a safe place to keep working.

Do I have to be technical to move my setup between tools?
No. The files are plain text in ordinary words. On the web versions you upload them into a project; on command line tools the folder sits on disk and gets read automatically. The technical part is not writing the files, it is knowing each tool's limits — file counts, size caps, filename rules — which is what the table above is for.

What do I actually lose when I switch?
You keep everything you wrote down. You lose everything the tool worked out about you on its own. As of 31 July 2026 no vendor offers a machine readable export of that second kind, and the two largest import features work by asking a rival product to write a prose summary that you copy across by hand.

How long does moving take?
Uploading takes minutes. Getting the same quality of output takes longer, because the limits differ. Budget a real afternoon for the first move and write down how long it took, so the second one is a known quantity rather than a fear.

Where you're running it, and how it connects

Honest ratings. Green means it reads the file as it is. Amber means you can get there with a workaround. Red means this one does not exist here.

Sit with that last row. No workaround closes it, and it is the reason the folder has a memory/decisions.md in it at all.

Related in this series

• Splitting one long set of instructions into separate job files — where the jobs/ folder came from.
• The list of jobs you must not hand over — why boundaries.md exists, and why it does not change when you switch.
• What a vendor's memory actually holds — the full breakdown behind the red row above.
• Running the whole thing for small business over the long haul — what to redo when models change, and how to hand the folder to someone else.

Already working from a terminal? The older technical write-ups here go deeper on file layout and loading order. They assume a command line; the material above does not.

The folder is yours. The model is rented. Check the lease occasionally.

Stay in the loop (no account signup)

This site does not ask you to create a product account. Free readers just leave an email—or follow where the build is posted.

No paywall on this article. No "sign up for access." If you only want one next step: use the email box at the bottom of the site, or follow on X if you prefer the timeline.

— hh, AI Workflow Pro