Winning in AI coding agents space

Table of Contents

Problem : If you are PM for an AI coding agent what strategy you will use to win?

Approach

In any domain if you are building a product, few things are important :

• Which customer segment are you targeting?
• How do you build a differentiated product experience for them?
• Do you have any (durable) competitive advantages?

Context

Broader trends in AI coding agent domain

• LLMs are good at code related tasks : generating code, debugging code (context), writing tests, etc.
• Building low-mid difficult software products is becoming more accessible to low-no coders. A lot of boilerplate development tasks are becoming conversational
• Evolution from copilot (ask me anything about software development) to agents (do software development work for me…)
• Low-mid complexity software can go creator economy way (publish videos, podcasts, etc → publish mini software)
• Overall Software Product Development is becoming more & more efficient!

What does these trends lead to :

• More non-tech people will hear “hat” of software builder and would bring ideas to life!
• More software will be built →
  • long tail → apps-for-1 or personalised apps, software for niche audience (earlier economically not viable but now is!), software for specific industry verticals
  • More efficiency means companies can use existing resources to build more…
• Professional Software Product Development teams will be smaller (Engg, PM, design) with big goals!

Domain - Software domain with coding look like…

Coding agents are an interesting space because LLMs are inherently good at coding and Software Engineering as an industry will disrupt for sure!

AI will seep into Software Product development at all stages….

• Scoping
• Product Development (PM, Engg, UX)
• Testing
• Managing Infrastructure
• Ops (delivery, efficiency, collaboration etc) : CI/CD, Program management, etc

User Segments

Criteria for creating segments : goal of building, type of software (web apps, mobile apps, marketplace apps, industry vertical specific apps)

• **Vibe coder building personal software (**Analysts , PMs, business folks,) : Minimal understanding of code but wants to build software for personal projects, showcasing ideas at office or to the world, efficiency apps for their productivity, etc. Want fast, cheap way to get ideas to life!
• Indie hackers building small-mid scale software products: Software engineers as they understand code but building interesting projects for themselves or even paid software professionally! Have to build sustainable income with software so need to acquire, retain customers in a profitable manner! have limited budget to spend!
• Professional software engineers working in companies building software products. Need agents for productivity and moving faster!
  • Size of company : SME, Mid-market, Enterprise
  • Industry vertical : More & more software should be built for non-digital niche spaces : Real estate, telecom, operations, manufacturing, tourism, etc

Picking a segment depends on a lot of factors, growth, your core competency (product vs sales, etc). Lets pick “vibe coders” which is growing fast and cheaper to build (less complex software)..Challenge is highly competitive!

How do you build a differentiated product experience?

• Dig deeper into their goals and segments and fulfil them better
• Leverage or build core competencies you have to gain an edge!
• Reliability (hard to build in AI products)

Vibe coders : Analysts, PMs, Business folks, etc

Goals :

• Bring their ideas to life! These folks have a lot of real world context and hence have so many ideas but were restricted by what was possible to build!
• What is possible? To what level they can go (Can the product match their high ceiling of imagination) (Games, mobiles, web apps, internal apps that can connect to various systems, etc)
• They want to showcase their ideas to the world as easily as possible!
• Code, testing, security, infrastructure,observability etc doesn’t matter and would rather be oblivious to it! Only outcome matters!
• Being able to manage change lifecycle : allowing to make changes, commit changes, rollback
• Cost conscious as these projects may not fetch any money to them!
• Finicky! Keeps moving from idea to idea (build fast/kill fast)
• Guide them towards building better - Help them assume the role of UX, PMs, marketer, sales, etc

Emotions :

• Empowerment & pride!! Products need to tap into emotions of empowerment (see what you can do now something like canva did for bad designers like me)
• The output needs to be good! Showcasing stuff with pride!

Key barriers in achieving above goals :

• Current coding agents don’t match their level of imagination (fall short!)
• Stability of current tools : The LLMs are prone to screw projects leading to people getting dejected
• Tools not handling : “Code, testing, security, infrastructure, observability and maintenance” -> make them redundant! Completely hide them!
• Tools can be costly sometimes!
• Keep collaboration like git etc are just added overhead!
• Focused only on building but what about allied functions -> how do I design better, how do understand & write requirements better, how do I market or sell my stuff better (AI powered marketing ,selling, videos, etc)

Leverage or build competitive advantage

- LLMs are commodity tools available to everyone!

• Can you do a “task” for your user segment “better” at cheaper “cost”

• AI competency stack for companies

  • (Table-stakes) Application layer optimisation : Can use existing LLMs to the best to solve customer needs (prompt engg., context management system, LLM + Tools, Evals, etc
  • Finetune for specific tasks : Can you take pre-trained LLMs and tune to your advantage for your tasks & data (fine-tuning, Reinforcement learning)
  • Model Inference Infrastructure : Can you host your own LLMs for cheaper (cost per inference token) & (faster → reduced latency) or partner with someone. Innovations like using more thinking tokens for complex tasks. Since this variable cost with AI coding agents companies which keep this in control grow much faster!
  • Pre-training special purpose small LLMs from scratch for your tasks!

Examples :

• Windsurf has built technology competency across all 4 https://codeium.com/blog/our-model-strategy. They have also built expertise in model inference infrastructure
• Bolt (stackblitz) built expertise in web containers (essentially web based developer infra) in a much more efficient manner! This helped them build “bolt.new” coding agent and keep it low latency to build web based AI coding agents and much efficient to scale (vs those are serving from cloud VMs)
• Replit was already building web based developer tools to democratise coding!

Existing Competitive advantages (7 powers)

• Previous knowledge or work can be leveraged (Replit, etc)
• Network effects, regulatory, ….scale, patent, distribution advantage…proprietary data!