According to the internet lately, AI agents are the next wave of innovative AI tech that’s going to make us more efficient, make our lives easier, take our jobs, or end the world. Okay, maybe not that last one, but still, the point is that there are a lot of opinions flying around about AI agents.
What I find interesting though, is that unlike large language models, there doesn’t seem to be any big name frontrunners in the AI agent race yet. Sure, there have been a few quiet releases here and there 1 2, but they’ve stayed mostly under the radar. OpenAI’s January release of the Operator model 3 is a rare exception, but that’s only because anything OpenAI puts out at this point catches some press. Despite its higher visibility though, in practice, Operator is too cost prohibitive for the average ChatGPT user to try. Unless you’ve got a spare $200 a month laying around, the only thing you’re going to have access to are more opinions.
But fret not my friend, because a new AI agent has entered the chat – Convergence. Or more specifically, Proxy, which is the name of Convergence’s AI agent.
After hearing all the buzz and finding out that it wouldn’t cost me $200 to try it, I decided to take it for a spin. But before I get into the fine details, let me first give you an overview of the two plans they offer and what they promise to deliver. Then I’ll share the results of three tasks that I ran using Proxy.
Convergence AI plans 💰
Convergence offers a free plan alongside a $20 per month Pro plan. The details they share up front only tell part of the story though. You’ll see why shortly. Nonetheless, they give you the gist of what you get with each option:

The Convergence AI promo video 🎥
The Convergence website has a two-minute-ish promotional video that highlights the benefits of using Proxy.
The benefits include:
- Multi-system and multi-instance operation: Proxy can navigate and work across multiple platforms, such as Salesforce, G Suite, Slack, Bamboo HR, Indeed, and Glassdoor. The video also mentions that “infinite numbers of Proxy” can work in multiple departments simultaneously.
- Personalized learning: It learns from your instructions “exactly how you would teach a new colleague,” adapting to your specific way of working. Once it learns how to complete a task (like creating an HR department in Bamboo), it remembers the skill for future use.
The video positions Proxy as a “generalist Twin” or “extension of you.” If you’d like to see it yourself, I’ve embedded it below:
Promo video versus real user experience 🧑🏼💻
Much like some of the other AI agent demos I’ve seen, my impression after watching the promo video left me feeling like this is indeed a game changer. However, that impression gradually spiraled downward when I actually signed up for a free account and put it to use. There is just no way anyone is going to take over the world using Proxy.
I ended up asking it to complete ten tasks over the course of two days (since the free account limits you to five requests per day).
The tasks that were fairly simple in nature and mainly just relied on completing a few Google searches were successful. However, more complex tasks that required Proxy to log into accounts required way too much of my human intervention and didn’t always yield good results.
For practical reasons, I’m not going to share all ten tasks here. Instead I will share three examples to give you an idea of what you can reasonably expect to get out of Proxy.
First task: log into my ebay account and search for camera lenses
After watching the video and seeing how seemingly easy it was for Proxy to log into various software tools to complete tasks, I wanted to see it for myself. Here’s the exact prompt I gave it:

It got to work right away.
However, despite giving it my login details, it very quickly ran into a roadblock.
You can see in the screenshot below.

In the steps it said that it was going to “securely enter [my] password to complete the sign-in process.” However, instead of actually doing that, it gave me the sign-in window and asked me to do it myself.
This was indicated by the orange dot next to the window, which signals that a manual intervention is needed.
Proxy asking for help
Anytime you encounter that orange dot while using Proxy means you have to tap on the Enter browser button to go into its browser and take control. The browser initially loads in a full-screen format, so you might have to magnify the screen to zoom in on the section you need to use.

In this example, after entering my password, I tapped Submit on the bottom right and this resumed Proxy’s attempt at completing the task.
At this point I was thinking to myself that while perhaps not maximally efficient, this one minor hindrance wasn’t so bad.
Unfortunately, my optimism was quickly crushed. Ebay flagged Proxy’s login attempt as “suspicious.” It then asked me to request a security verification code via SMS.
The most frustrating aspect of this, was that Proxy wasn’t contextually intelligent enough to simply click on the Text me button. It once again asked me to manually intervene to literally just click a button:

Unfortunately, ebay still wasn’t convinced and I also had to complete a manual captcha challenge:

And after all that, I had to manually sign in again:

This wasn’t the time-saving efficiency that was promised to me in the promo video.
Results of first task
Despite the initial roadblocks I encountered with logging into my account, in the end Proxy was able to sift through ebay and come up with a list of five camera lenses for me, including links to each listing:
In Proxy’s defense, a lot of those login hurdles were simply a result of security processes that most websites have.
But therein lies the somewhat misleading marketing promise.
As a regular everyday user, you will likely encounter these login issues on almost all of your accounts if you ask Proxy to log into them. This is because they will all flag Proxy as the bot (excuse me, “AI agent”) that it is.
At least initially.
It’s possible that after having you jump through all these manual hoops on an initial Proxy request, that some websites will green flag subsequent attempts to log in.
Having said that, I wouldn’t get my hopes up. The reason is because it appears that Proxy shuffles around IP addresses when using its browser. I noticed this when I got emails from ebay that matched the manual intervention requests by Proxy:

In short, I wouldn’t bet on Convergence AI saving you time on any tasks that require it to log into accounts.
Second task: search the web for remote content writer job openings
For the next request, I wanted to see how Convergence might help someone looking for a job. Here was my exact prompt:

One thing I didn’t share with you in the first example, is that you can actually observe Proxy doing its work in a sidebar window. This opens up by default when you first use it, but you can hide it or show it by using the toggle button shown above.
This feature is somewhat interesting the first time you use Convergence, but the novelty wears off very quickly. After all, the whole point of using it is so you can do other work while it’s doing the task you asked it to.
Results of second task
Compared to the ebay request, the results here were much better. I didn’t have to do any kind of manual intervention. Proxy simply searched around the web and compiled a list that matched my criteria:

Versus Perplexity and Google
For a thorough analysis, I decided to check how Proxy’s results would compare to Perplexity and Google.
I began by giving Perplexity the identical prompt that I gave to Proxy:

Although Perplexity was infinitely faster at producing its list (it is a search engine after all), in my opinion, the quality of the Convergence list was much better.
For the Google comparison I shortened the query to make it more appropriate for a standard Google search:

Convergence got the winning nod against Google as well.
Overall, for this kind of relatively simple task, I think using Proxy makes sense. I could see myself coming back to it again for similar requests. However, there is still a caveat here, and it has to do with the size of the query.
The third and final example illustrates this well:
Third task: search the websites of 25 companies for remote content writer job openings
As you just saw, tasks that don’t require Proxy to log into any accounts and only search the web tend to produce good results.
But there are limitations to how much you can ask it to do. For example, I gave Proxy a similar prompt about finding jobs, but this time I gave it a list of 25 specific companies that I wanted it to check.
For context, I got this list from a LinkedIn post that stated that these specific companies were currently hiring writers. In other words, I didn’t send Proxy on a random goose chase. The list was intentionally chosen to make it easier for Proxy to complete the task.
Results of third task
At first, everything seemed like it was going well:
But then, after researching the seventh company out of the full list of 25, it decided that it had enough:

And that was it. There was no message about it not being able to continue. It just marked it as complete and pretended like nothing happened.
I tried this with several other queries and had similar non-results every time. With one particular query, it did initially mark it with a red flag after reaching 39 steps into the process, but bizarrely, the next day the red flag disappeared and it was added to my list of “completed” tasks.

If you plan on trying Proxy yourself, the two takeaways here are:
- 👉 Limit your requests to a manageable size.
- 👉 Pay close attention to the steps and ensure your request has been completed in full.
My advice is to leave your queries open-ended and let Convergence decide on its own when it wants to stop searching. If it asks you a follow-up question with a request for details, keep your expectations small (i.e., up to five of something).
Other notable features ⚡️
Before wrapping this up, I want to briefly highlight two other features that Convergence offers:
These are the ability to set up recurring tasks, and to use templates.
They are exactly what they sound like. After you complete a one-off task, Convergence gives you the option to convert it into an automated (recurring) task or to make a template out of it.

The main difference between the two is that with the automated tasks, you set up a schedule for the task to occur on whatever basis you decide:

With templates, it sets the prompt aside for you to reuse in the future whenever you wish.
You can also share the template with the public Template Marketplace, as well as search templates created by others:

My final verdict ⚖️
After securing $12 million USD in pre-seed funding last year 4, the hype around Convergence AI was real. The company released their Proxy agent about a month after that funding round 5. Even though I was slightly late to the party by a few weeks, I was excited to try it. Unfortunately, I was left mostly disappointed. Aside from using it for simple search queries, I don’t see myself coming back to it for more complex needs until they release a few more iterations.
Perhaps the $20 a month plan is significantly better, but the free version left much to be desired. It’s similar to how I felt about OpenAI’s Sora after trying it a few months ago – cool but not quite there yet.
I have a gut feeling that most of 2025’s AI releases are going to be categorized as cool but not quite there yet, but perhaps something will come along to pleasantly surprise me. Until then, I’ll keep watching in amusement as AI company CEOs predict AI surpassing humans at almost everything inside of three years. 6
Sure…I’ll believe it when I see test it.
Have you tried Convergence AI’s Proxy or any other AI agents? What did you think? Let me know in the comments below.
- https://github.com/reworkd/AgentGPT ↩︎
- https://github.com/Significant-Gravitas/AutoGPT ↩︎
- https://openai.com/index/introducing-operator/ ↩︎
- https://tech.eu/2024/09/27/convergence-secures-12m-pre-seed-for-personalised-ai-agent/ ↩︎
- https://www.linkedin.com/posts/convergencelabs_were-excited-to-announce-that-proxy-europe-activity-7289567931728896002-9SN3 ↩︎
- https://arstechnica.com/ai/2025/01/anthropic-chief-says-ai-could-surpass-almost-all-humans-at-almost-everything-shortly-after-2027/ ↩︎
…
Don’t forget to join our crash course on speeding up your WordPress site. Learn more below:
A comment about Proxy giving up part way through: I’ve been having similar problems with ChatGPT and asking it to process text documents. It keeps starting and giving up part way through (and thinks it’s finished). It seems to be the length that’s a problem – give it 1000 words, and it’s OK; give it more, and it can’t cope. As with yours, no awareness from the LLM even when explicitly asked, and also when the LLM thinks and promises it can do more.
Thanks for reading Neil and for your comment. I agree. There is a real disconnect between the hype that’s being sold by the general “AI tool industry” and the reality on the ground. I understand they are doing this to raise capital, but I think it’s also having some disastrous real-world consequences because “AI is coming for our jobs” is being taken literally by some people in decision-making roles who aren’t bothering to do simple test runs of their own or even bothering to ask their staff what their experiences with these tools has been like. They are firing their humans and then attempting to use AI to replace the function of those humans…and everyone is losing. The workers because they are out of a job, the end users/consumers because they are being fed slop by the AI, and even the companies themselves who are either getting bad PR from it or just a general consumer backlash. We’ve seen this already in countless examples from content to SaaS. The bigger companies have simply taken the hits and re-hired some of their humans back but the whole thing has been a mess. AI tools have been great, but for the most… Read more »
The first issues with logging in is the fault of the user. It intentionally hands off login tasks for various reasons including for security. This is by design. eBay also did what it should have to limit the automated bot activity.
AI Agents can do some amazing things, but it still requires the user and other applications it interfaces with to support them. It can be helpful for users to be able to automate processes, but many malicious bots automate harmful activity which websites still need to limit.
Thanks for your comment Jason. Yes, I stated in the article that this is by design and that ebay’s actions were normal security precautions.
“In Proxy’s defense, a lot of those login hurdles were simply a result of security processes that most websites have.”
Is there something you would have done differently to avoid the logging in issues? I’m asking because you wrote “The first issues with logging in is the fault of the user.” so I’m curious if you have a tactic that readers might benefit from. Thanks again.
Thank you for the reply, Martin,
I’m sorry if I came off too negative. My main complaint was that it appeared that you were suggesting that the login issues were mainly the AI Agent’s fault, rather than fundamental hurdles of our current internet infrastructure; but you did try to explain that in the article.
I believe it’s a good thing that Convergence doesn’t yet try to store and use credentials using Agentic AI yet due to security and other issues. It looks like there are some cool new technologies which will help with this though, such as Verifiable Credentials (VCs); which could offer a more secure and flexible way for AI agents to handle authentication. Also, stuff like ephemeral authentication could help by creating short-lived, task-specific credentials.
Thanks again!
When a complex task is encountered, it will display ”
I reached my maximum number of working steps and could not finish my work.”🤣