Taking Convergence AI's "Proxy" for a Spin: Your Agentic Assistant?

The buzz around AI agents – autonomous systems that can understand tasks and execute them across different applications – is growing louder. We're moving beyond simple chatbots to envision AI that can truly act on our behalf. Enter Proxy by Convergence (found at convergence.ai), a platform positioning itself as "Your AI assistant for your daily tasks." Is it the agentic future we've been waiting for? We took a first look based on a recent walkthrough video.

What is Proxy? First Impressions

Proxy presents itself as an AI designed to offload and automate repetitive tasks. The landing page showcases examples ranging from information gathering (summarizing pull requests, news articles, or academic papers) to taking real-world actions (buying gifts on Amazon, booking restaurants, buying groceries).

The core idea follows a simple three-step process:

1. Ask Proxy to complete tasks.

2. Proxy gets back to you with the results.

3. Schedule Proxy to repeat tasks daily (automation).

The platform is backed by notable names like Balderton, Salesforce Ventures, and Shopify Ventures, suggesting significant investment in this vision.

Upon logging in (the reviewer used the free tier, which offers 5 sessions per day and 1 automation), the user is greeted with a clean dashboard. A central input field prompts, "What online tasks or research would you like Proxy to do for you?" Below this, pre-defined task templates are organized into categories like Business, Marketing, Research, News, Sales, DevOps, and Fun, providing easy starting points.

Putting Proxy to the Test: Two Tasks

The review explored two main tasks:

1. Brand Analysis:
Using the "Check News about Brand" template under Marketing, the reviewer asked Proxy to perform a quick brand analysis on "OpenAI," limiting results to the past week.

• The Process: Proxy initiated the task, and interestingly, a "Proxy's View" panel appeared, showing a live-like view of a browser window where the AI was performing actions. It opened Google, searched for "OpenAI" news within the specified timeframe, identified relevant articles (like one from Reuters), and began extracting information.

• The Result: After a short wait, Proxy delivered a well-structured "Brand Analysis Summary for OpenAI." It highlighted recent developments like the new Responses API and Agents SDK (which was indeed recent news), discussed the competitive landscape, and summarized sentiment (mostly positive reception but concerns about cost/customization).

• Impression: The reviewer found this capability "pretty cool" and the results "pretty up-to-date." The ability to see the AI "working" in the browser view was a neat touch, adding transparency.

2. Planning a Night Out:
Switching gears, the reviewer gave a more open-ended prompt: "Can you help me plan a great night in New York City? I want to spend $100 dollars and have dinner at a vegan restaurant and maybe do something with the arts that is cheap?"

• The Process: Proxy started by asking for clarification: "Could you please specify which day or time you're planning for...?" The reviewer responded, "Sunday early afternoon." Proxy then used its browser view again, searching Google Maps/Search for "NYC vegan restaurant dinner $100 budget," exploring specific restaurant options (like Ladybird), searching for free/cheap art events on "The Skint" website for Sunday, and compiling an itinerary.

• The Hiccup (CAPTCHA!): During the process, Proxy encountered a CAPTCHA on "The Skint" website. It couldn't solve it automatically and prompted the reviewer to take control of the browser temporarily via an "Enter Browser" button to solve the image challenge (selecting crosswalks, then bicycles). The reviewer noted this was "wild" and "kind of fun" but also highlighted it as "clunky" and "slow," indicating a current limitation of the AI agent.

• The Result: Despite the CAPTCHA interruption (which the reviewer eventually bypassed manually within the AI's browser view), Proxy successfully completed the task. It proposed an itinerary: Dinner at Ladybird (providing location, price range estimate, hours, and recommendation) followed by an Art Activity: Visit The Metropolitan Museum of Art (noting admission is "Pay-as-you-wish" on Sundays and providing location/hours).

• Impression: While the CAPTCHA interaction was a definite "bump," the final result successfully addressed the user's request within the specified constraints.

Overall Thoughts: Promising but Early Days

Proxy by Convergence feels like a tangible step towards the agentic AI future, reminiscent of concepts like OpenAI's (now defunct) Operator. The ability to define tasks, leverage templates, watch the AI work through a browser view, and receive structured results is compelling. The free tier makes it accessible for experimentation.

However, it's clearly still evolving. The tasks take a noticeable amount of time to complete, and the inability to handle CAPTCHAs independently, requiring manual user intervention within its own browser session, shows it's still "rough around the edges." The need to manually solve the CAPTCHA felt particularly cumbersome.

Key Takeaways:

• Pros: Agentic approach, task templates, transparent browser view, free tier for testing, potential for complex workflow automation.

• Cons: Can be slow, struggles with CAPTCHAs (requiring clunky user intervention), potential security concerns for deeper integrations (like calendar access, though not tested here).

Should You Try Proxy?

If you're interested in the potential of AI agents and want to experiment with automating information gathering or simple web-based tasks, Proxy by Convergence is definitely worth checking out, especially with its free tier. It provides a fascinating glimpse into how AI might soon handle multi-step processes for us. Just be prepared for a few bumps and a bit of a wait as it works its magic – it's a promising tool, but still on its journey.

You can explore it yourself at convergence.ai.