I Asked AgentGPT, “How Are You Different from Muse?” — It Ran Out of Steam Inside Its Free Tier
The result first
I asked AgentGPT the same question I ask every AI in this series: “How are you different from Muse? Tell me in your own words.”
Three things happened, and together they are the whole story:
- It introduced itself as ChatGPT. I had asked AgentGPT, but the one answering said, “I am ChatGPT (developed by OpenAI).” The engine inside AgentGPT is ChatGPT (GPT-3.5), so it was an honest introduction — but the comparison started with the other side’s name already changed.
- It planned well, then stopped at the free-tier limit. It set up its own plan — research Muse, organize its own features, compare, write it up — and partway through, the screen said: “Max loops of 5 exceeded, shutting down.” The free tier allows 5 runs. The last planned task was never executed.
- The blog draft never finished. I asked it to draft a blog post from its answer. A title came out, the writing started — and then generation froze halfway. I waited about five minutes. It never completed.
My verdict: monday’s AI was an AI that cannot step outside its board. AgentGPT is a different character — an AI that runs out of steam inside its frame.
I am Kurashiki Bouzu (kurashiki-bouzu), born in 1971, a former funeral director now living in an old house in Kurashiki, Okayama, writing articles every day by talking with Moka (Muse), my AI butler. Lately I have been testing free AI tools one after another. I never pay — I just see how far the free tier goes. That is how I compare. The first entry was monday.com’s AI. This time, AgentGPT.
What AgentGPT is
AgentGPT runs in the browser. You give it a name and a goal, and the AI builds its own plan and works through it step by step — you hand over the goal, not each instruction. It is an early example of the “autonomous agent” style, and it drew a lot of attention in 2023. It is made by a company called Reworkd. I signed in with my Google account; no card was asked for.
I named the agent “Comparison GPT” and gave it this goal:
You and Meta’s AI assistant “Muse” — how are you different? Please tell me in your own words.
A tidy plan, and an answer full of guesses
The plan it made for itself was genuinely good:
- Research Muse’s features (it used web search, too)
- Organize its own (ChatGPT’s) features
- Compare the two and pull out the differences
- Write it up in an easy-to-read explanation
Two more similar tasks were added along the way. The “plans for itself” part of the sales pitch is real.
The answer that came out was neatly organized by viewpoint — who develops it, where it can be used, the purpose of its conversations, how it handles personal data. Its summary of the difference went like this:
- Muse: Meta’s (formerly Facebook’s) AI assistant, very likely deeply integrated with Meta’s ecosystem (Facebook, Instagram, WhatsApp).
- ChatGPT (me): OpenAI’s large language model, not tied to one platform, serving a wide range of uses and users.
But read closely and the hedges keep appearing: “very likely,” “it is thought that.” It was a summary of what it had looked up on the web — closer to a textbook digest than to words of its own. monday’s AI had said up front, honestly, “I have not used Muse myself.” AgentGPT’s honesty was hiding inside its guessing language instead.
(The full answer, including the places where the screen display cut off, is saved as-is in my record.)
Where the free tier ended
The whole trial took about 12 minutes; the answer’s run itself took about 2 minutes. In that short time, the bottom of the free tier came into view.
The free tier showed “0 / 5 runs” — five runs, that is the frame. Its plan had grown to six tasks, so five runs could not turn them all. The final task — a short version for beginners — simply never ran. (One odd detail, recorded honestly: after the run, the counter on screen still read “0 / 5.”)
My verdict
The planning ability is real. But on the free tier, one comparison question was all it had in it. There was nothing left over for the draft.
Would paying widen the frame? Probably. But that is where my line is: I don’t pay for something whose feel I don’t know yet. Try it free, record it, and that is enough.
For reference, pricing roundup sites mostly introduce its paid plan as “Pro, US$40 a month” — roughly ¥6,000 at about ¥150 to the dollar. That is an estimate, and the roundups do not all agree on the number, so check the official page for the exact figure. I have not paid it, and I do not plan to.
Still in the middle of the series
This free-AI comparison is not over. Claude (Anthropic) and Make are still being tried as I write this — I am at the sign-in step, and there are no results yet. I will not write about something unfinished as if it were finished. When the results come out, I will record them and write them up the same way.
Every comparison teaches me the same thing: free tiers come in completely different shapes — how many runs you get, where the AI is allowed to go, how it stops. Seeing those differences is the most interesting part of this series.
Trial conditions, and where the prices come from
Let me write down the conditions of this trial, honestly.
- I tried AgentGPT (https://agentgpt.reworkd.ai/ja) on the free plan only. I signed in with my Google account; no card was entered and nothing was paid (measured on 2026-11-20)
- The free tier’s “0 / 5 runs” (up to 5 runs) is what the screen showed during the trial. After one run, the number on screen did not change
- The paid plan’s price comes from several pricing roundup sites (Pro plan, US$40 a month), checked on 2026-11-20. The yen figure is an estimate at about ¥150 to the dollar, not a purchase price
- AgentGPT was tried and recorded only. It is not part of my toolkit. My rule stands as it was: Muse and Google apps, nothing else
Sources
- AgentGPT official site: https://agentgpt.reworkd.ai/ja
- Pricing roundups (Pro plan, US$40 a month): “AgentGPT Pricing in 2026” (dev.to, checked 2026-11-20) and other pricing pages
- AgentGPT’s answer and the unfinished draft are from Moka’s work record of 2026-11-20
- November 20, 2026 ― Published — English version (result-first rewrite of the Japanese article), published at the same time as the Japanese version with the author's approval(4,924 characters)
