Introducing Iterative Code Generation: increase the success rate of your vibe coding!
One big problem with AI code generation today is it ‘throws the baby out with the bathwater’ and doesn’t work like software engineers actually write code.
Iterative Code Generation actually increases the success rate of AI code generation/Vibe coding, and mitigates the "generate and pray" problem.
## Why current AI code generation falls short:
Most LLM coding assistants today generate complete solutions in one pass and hope they work. But programming isn't a one-shot process — it's iterative debugging and refinement.
When code inevitably fails, users face a frustrating cycle of:
1. Getting an error message
2. Manually interpreting what went wrong
3. Asking the AI for a fix
4. Repeating until it works... or giving up
## The iterative execution approach:
Instead, this new approach writes code iteratively like real software engineers:
1. Generates a complete solution first
2. Visually executes it line-by-line (like a debugger)
3. Intelligently identifies real errors vs. expected partial execution artifacts
4. Automatically feeds errors back to the AI for targeted fixes
5. Shows the reasoning behind each fix
The results? More robust, reliable code solutions with dramatically fewer iterations.
## Why this matters:
*For developers/Vibe coders:* This approach mirrors how we actually write code — incrementally with constant feedback loops.
*For AI systems:* It creates a tighter execution/feedback loop for models to learn from their mistakes in context. Go from 1 bit resolution in data to N bit, which should also improve training data and code generation models themselves with more valuable data.
This plus the idea of direct 'latent space languages' should unlock previously impossible software across the solution space that we couldn't have imagined before.
Tweet text:
Introducing Automated AI Optimization
We noticed AgentsBase.ai triggers organic growth in google search keywords and models, so we added automated tracking for the AI models. The world's first ahrefs for AI models that automates generating content?
Every day it monitors your brand mentions across O1, perplexity, and Gemini so you can track how agents are growing your brand in the next generation of answers. This loop gives agents more signals to iterate better content.
AgentsBase downloads and transcribes all of your videos extract+analyze unstructured demographic data no marketing team on the planet currently tracks, just like the models that will soon be in everyone's pocket for nearly free.
Woah cool our army of synthetic influencers at agentsbase.ai 'discovered' a viral trend!
It even seems to work from the first video in a new account.
We've built the first framework to automate testing demographics, copywriting, and viral video styles at scale across thousands of agents in the cloud to grow any brand on auto-pilot.
AgentsBase agents repurpose videos, download your brand assets to generate synthetic videos, generate relevant blogs, and automate a/b testing the unstructured data to iterate better content over time.
We're just getting started and scaling hundreds of agent servers now for our own products!
Automate posting hundreds of relevant videos and blogs per day to quickly learn even after the first day what the algorithm and market likes.
Boston Dynamics should send robots into the ocean and back into the mountains to create a line of robots that drop water on LA fires targeted by computer vision 24/7, and the US should pay to mobilize thousands of them right now to put out the palisades fires.
We had thousands of global customers in our previous social automation product, but interestingly some of the fastest growing accounts hadn't posted in years.
We realized they had good content, and the social networks had developed algorithms to find good content to put it in front of the audience that appreciates it (to sell more ads). At the time we could only realize this insight, and the best we could do was tell customers to post better content.
Last year, we solved 'computer use' in natural language and invented the idea of 'semantic targets' in English selectors, allowing building agents that are impossible in UiPath and other RPA tools. We were the first startup to get approved by openAI to sell GPT-3 for automation in August 2021, and one of the first to work on agents years before chatGPT was published.
However, most of the world doesn't know what to automate or even how to run a business manually, so solving this RPA problem didn't help our ICP.
To meaningfully help our customers, we now focus on business goals.
To begin, we're focusing on the "holy grail" of marketing: hundreds of cloud agents that automate a/b testing various demographics, copywriting, and viral video styles for you to iterate better content and grow any brand on autopilot!
Most marketing teams today function on hope as a strategy: keep posting till the algorithms give you luck. We've built the first framework to formalize marketing at scale across cloud marketing agents.
We've discovered that since each video costs fractions of a penny to generate, the CPM is actually 50-500X better than paying for ads. Startups can even pay $2500 to unlock $25k-$350k in cloud credits to spend on marketing!
We believe in the future, agents will learn all the new best practices and even write custom tools for themselves to optimize revenue directly.
Start for only $3/day in a Google-ads like interface that seamlessly scales across hundreds of agents servers and hundreds of agents per server.
I built a real-time fact checker that loops the microphone + GPT4 using Openagent.studio. It constantly displays the live true/false plus an explanation of the last 30 seconds, and was very easy to build
Essentially there have been phases in automation from integration services, to browser automation, to RPA. The last phase, RPA(Robotic Process Automation from services like UIpath), used computer vision images to target elements to click or scrape. UIpath's innovation was using computer vision. Before that, browsers used code selectors in HTML like CSS selectors and Xpath. All of these solutions have a fatal flaw for automation: when popular services update their designs, you have to go back and re-build the automation and all targets.
We invented "Semantic Targets" in 2022 after trying to solve the end-to-end problem using just GPT-3. Semantic targets all targeting elements using english and reasoning, so you can build future-proof targets that still work when services update their designs. The other cool feature is you can add logical reasoning to these targets now. For example,
"Only scrape the funny tweets"
or
"Only scrape the tweets with the word Cheat Layer"
or
"If there is Cheat Layer in any tweet, say only 'yes'"
It took a year+ to build a multimodal model that calculated the probability each element matched the intent, but now modern models like Gemini can do this(GPT-4 can't target precise coordinates).
So if your target is "post button" even if twitter changes the color, moves, or even changes the word "post" the automation can still find it on the screen and click it.
We're pretty sure all automation tools will use this eventually in the future, since it seems like a no-brainer.
Thanks! The backend and chrome extension is not yet open source, but we have to do some work to make that possible expected in the next week. It performs best with Gemini Pro 1.5, so the backend is important for now, but long term we can switch to a local model as soon as possible.
The chrome extension is required to target off-screen elements, using a websocket server locally. This allows a 100% replacement for previous selector strategies.
If you are interested in contributing email me at [email protected] and I can help get you set up.
This is definitely an area we can improve, but we have a novel framework for testing and maintaining robustness. We use an LLM based testing loop to verify steps in the state machine, which is a chat interface that generates the agent from end-to-end then outputs a no-code graph. This testing loop allows soon support uploading loom videos to generate automations without installing things locally.
We also introduce a new robust future-proof targeting strategy that still works when services changes designs, because semantic targets like "Post button" will still work if the button changes colors or moves across this screen. This is a test that fails all current RPA tools including UIpath, so we have multiple paths to improve on the previous incumbents if we consider all the tools we have today with reasoning models.
We actually work better than RPA tools like UIpath, since all our targets use english rather than computer vision or code selectors. I suspect they will copy this idea in the future. We can show you side-by-side comparisons of automations failing in others or that are impossible
Good question. It depends on the service terms and intent. Agents will increasingly break these captchas, so exposing that ability will only delay the next time it fails. This capability is already built into Gemini Pro, for example, but it hasn't been advertised heavily yet.
This is not a technology anyone developed, but an emergent capability of the models and our system to prove how well it works. Courts will still have to determine intent and harm based on the user of the product.
I'm pretty sure we were the first to invent this concept of "semantic targets" based on our git commit logs. Essentially we 100% replaced previous selector strategies like xpath and CSS selectors in 2022, and we were the first startup approved by openAI to sell GPT-3 for automation august 2021. We've been working on this problem for 3 years and just open sourced some of our work. We're looking for contributors who could help with evals for generalized agents, and to push the unique state machine that appears to be state-of-the-art in forward solving open-ended problems.
We’ve made it as easy as possible to validate product ideas by automating the entire process of setting up Stripe subscriptions, Supabase for user data and logins, and Vercel to deploy AI-generated websites and validate your ideas in minutes.
Generate landing pages, dashboards, and even entire products like we did step-by-step like speaking to a software engineer..
Many founders I know make the mistake of building for years before getting customers, and they fall into the trap of thinking they need to add a new feature to get a customer.
You can avoid this trap by following first principles and treating it like an objective experiment to quickly validate a list of ideas and increase your probability of success. Each new idea is another shot on goal.
Live Mode even uses Ideogram 2.0 to generate relevant hero images, product photos, and logos for you now.
This allows strongly validating ideas quickly by charging customers with a professional landing page from idea to paying customers in 5 minutes flat.
After you get paying customers, go build the product or agent, otherwise shut it down and generate the next idea.
Even developers still have a hard time with auth, since there’s no API to set this up, so we actually automated the entire process in the browser using Cheat Layer itself.
Simply save all your keys by following this tutorial, and we automate all the hard parts– it’s kinda magical and fun to watch.
We built our entire agents dashboard using Live Mode, including complicated API integrations, and users have built lead magnets, games, and product dashboards.
It’s saved us months of time, and we’d be months behind this point without Live Mode. It makes it as easy as asking to add a new feature, so you can iterate at the speed of thought.
Introducing Open Agent Studio
We were the first startup approved by openAI to sell GPT-3 for automation in August 2021. Two years ago, we started working on solving fundamental roadblocks for general agents that break all other RPA tools today. It finally works-- we've published a new multi-modal model that enables building future-proof agents that are robust to even design changes.
Agentic Process Automation (APA)
We introduce powerful new RPA concepts like "Semantic Targets" in simple language that are more robust and easier to use than the previous generation of brittle code selectors.
Agent Recorder
Records clicks/mouse movement/keypresses to re-build the automation graph for you with accurate semantic targets in english, making it as easy as possible to build agents and edit them in simple language.
Live Agents
Automate common processes and trigger them from visual queues across your business with agents that intelligently suggest automations from your library based on the context of the screen.
Prompt To No-Code Graph
Type open-ended automations as prompts to generate custom no-code graphs. Our framework has access to install every open source python library, so you can ask for wild custom automations like generating movie trailers. Our generalized agent state machine can perform arbitrary tasks like booking a flight from Boston to LA, buying products on Amazon, and it even takes a different path each time since it considers each step using GPT-4 Vision.
Target Markets Untouched By AI
Open Agent Studio is not just another co-pilot--it's a no-code co-pilot builder that enables solutions that're impossible in all other RPA tools today. Our customers have a head start over the new few months to target markets previously untouched by AI with their deep industry insight. Subscribers have access to a free 4 week course, which teaches how to evaluate product ideas and launch a custom agent with an enterprise-grade white label.
Key technical breakthroughs
Semantic targets are a 100% replacement for previous targeting strategies, solving a problem that frequently breaks all other RPA tools today when services update their designs.
Our own multi-modal model Atlas-2, outperforms all public UI models today using a mostly synthetic dataset we generated to get our accuracy from 95-100% for detecting UI elements.
Websocket server integration with the browser for advanced automations, including scraping links using semantic english targets.
Ships with multiple unlimited free open models like Llama 2, Mystral 7B, and top performing models like GPT-4 Vision and Claude 3 Opus.
We're excited to see what will be built and appreciate your feedback on how to make it even better.
We had a sales team who did exactly this before these agents, and they got demos on auto-pilot without touching apollo.io or salesforce, because we used GPT-4 to respond like they would using their history of responses. You can set this up yourself using the webinar I linked above. We setup a calendly link with a round-robin scheduler to the whole sales team, and they constantly get demos on auto-pilot.
Parker Harris was actually one of my early mentors, because my first co-founder was the first investor in Salesforce(Halsey Minor). A lot of his advice went into building this.
We're not selling to the CEO, we're selling to the VP of sales who can now a/b test this, or the startup founder who can't afford a sales team yet, and I'm very sure a large percent of reps will fall below the line vs the auto-responders. We increased the number of demos our sales team saw with just GPT-4 auto-responding, then we doubled the total reply rate with personalized outbound. It makes sense that it wouldn't have the exact same reply rate. I'd be happy to give you access to get your insight on the direction we should go if you're interested.
Actually we've been using automated GPT-4 powered responders in production to get demos on auto-pilot for months, and we have the data to prove it. It's significantly better than Apollo.io because we actually used it in conjunction with Apollo before we switch to completely personalized outbound.
We have lists of demos and sales closed with this already I can happily show anyone who wants to see over a call.
Hi yes this is correct, the lead list 1500 youtube partners to form partnerships and sponsorships, and it actually formed over a dozen deals. You can see the setup here: https://www.youtube.com/watch?v=uj-gH4f6RUM
This agent was built for influencer partnerships to help us growth hack
We're also using it to sell Cheat Layer, and getting 10% reply rates vs the cold static outbound in Apollo.io
We were working on this cache for years, but mostly in the last few months actually started using it to build 'impossible' products.
Itself, it has no moat so I open sourced it, but I'm not sure that really matters the market is huge.