Introduction to AI Coding agents
AI Coding Agents have proved to be a very successful use case for Machine Learning models. These coding agents are very useful because they can understand a question and spit out code very quickly. In most cases the code is decent boiler plate code… the rough edges and corner cases still need to be handled out by someone familiar with software. Still, its a great advance since writing boiler plate code can take 30-40% of a software engineer’s time.
AI Coding agents have made major progress over the last one year. As an example, Mark Zuckerberg recently said that Meta will have a mid-level engineer alike coding agent by the end of this year. A mid-level engineer at Meta typically earns $500000, so that is the value of economic displacement that AI Coding agents can bring.
In this post, we will explore how one can easily use AI coding agents to build near perfect code. There are many AI coding agents in the market… but the simplest use case is to directly query a model (such as DeepSeek-R1) provided by a cloud provider (such as Sambanova). For this exercise, we evaluated the code quality from multiple providers. Specifically, we evaluated Blackbox.ai, Perplexity.ai, and Sambanova. We limited ourselves to those providers who provide free versions of their offerings… implying, we didn’t spend a cent to generate this report.
The detailed configuration settings of the cloud based offerings we evaluated are:
Blackbox.ai: Used the “Deep Research” and “Beast Modes” in addition to specifying the agent chosen as “DeepSeekR1”
Perplexity.ai Used the Deep Research Mode
DeepSeek-R1 offered by Sambanova.
Note that the first two offerings allowed us to upload a text file which was very valuable to us for providing the list of animation strings supported by Vector, which we have publically available in our course git repository. At the time of writing this draft, Sambanova doesn’t allow one to attach a file to the query. Blackbox.ai and Perplexity.ai allow this feature.
The exercise we tried is to ask each agent to write code to generate a customized animation sequence for Vector robot using the Vector Python SDK.
Trying out Perplexity.ai
Let’s take a look at the specific case of how to use perplexity.ai to make a Vector robot dance with an animation sequence. The process is very similar for all the other providers. The steps are as follows:
Login at Perplexity.ai with your Google account (Or create a different login if required).
In the drop down box of query agents, you would want to choose “Deep Research”. At the time of writing this article, Perplexity.ai allows 5 FREE user queries per day on Deep Research.
Now, in the message box, write the following prompt: “I have an Anki Vector robot. I want to make the robot dance in a series of animations. Please write a program which can connect to my Vector robot using the Vector Python SDK, and accomplish the above. I would like you to use existing animations as well as create new animations. You can create a new animation by mixing two or more animations, and keeping the head moving from one animation, and having the lift moving from another animation. Try to get a syntactically correct program.”.
Here is a screenshot from Perplexity
Remember that this is just a sample prompt. Feel free to play with the prompt, experiment, and try different options such as other models. It is important to play with prompts to get a good result… this field is known as prompt engineering.
I needed to work on followup prompts. The initial code returned was good, but it had the animation strings incorrect. I then provided Perplexity with a file containing all animations and asked it to choose from the provided list. The code was still picking up some incorrect strings. I the provided Perplexity with a list of strings in the query. The code returned was neat and clean. Here is a thread of my Perplexity queries. Here is the Python code that got generated for me.