AI First Software Architecture

voxel crystal landscape

Edgar has been a passion project of mine for the past year, and I’m excited to announce that PSL is taking it to the next step and incorporating Picco Systems. Edgar is now Picco. We’re seeking business and technical talent to transform the raw potential of an AI-first software architecture into solutions for 100x more business use-cases. See the links at the end to apply for a position.

In the context of business software, we define AGI as the dynamic assembly of code, or ‘skills,’ to handle customer requests, or ‘tasks.’ A skill can range from a simple five-line API call to a complex assembly of prompt chains and other skills. We are currently ‘coaching’ the Picco system to create skills that accelerate the sales processes of PSL portfolio companies.

What is an AI-first software architecture?

In the past we had design patterns like “Model View Controller” web sites, or “mobile-first” apps. The power of LLMs is so great, we should re-organize our code to maximize how we use them.

1/ Functional Programming model: Everything’s a skill. I used to be an Object-Oriented programmer, drawn to how relatable objects are and how the code resembles English. But a graph of individual functions is much easier for an LLM to reason about.

In Functional Programming, each function is self-descriptive, with clear inputs and side effects. We use clean, well-documented skill templates. We chose TypeScript for its concise syntax and well-understood asynchronous semantics. The call graph (an example is below) becomes the agent’s skill tree.

Function Call graph or "Skill Tree"

2/ Skill Library: Similar to software modules, our Skills Library contains trees of skills (functions) in various areas of expertise. These serve as examples for generating more code and as building blocks to build upon. Selecting the right skill is a critical piece of the AGI puzzle (see Incremental AGI). Currently, Picco uses a simple in-memory vector store and, depending on the context, invokes an LLM to select the skill to invoke, or as an example for another generation, or to build upon.

3/ Extensive Prompt-chain Skills: Certain types of programmatic reasoning, previously unachievable, can now be accomplished using an LLM (summarizing text, writing, answering questions, subjective scoring, filling the blanks). Although there’s nothing particularly unique to Picco about these prompt chains, they are revolutionizing software development everywhere. In an AI-first architecture, they are integral to the software’s operation. Any robust prompt chain software must include tooling for switching models, maintaining test cases, estimating costs, and logging prompts and completions.

4/ Self-Improving: Picco’s “writeSkill” and “reviewSkill” are themselves skills that generate new and improve existing skills. While we use a traditional MVC framework for our website, adopting this model means that, wherever possible, everything is implemented as a skill.

What is Picco not? Today, Picco does not do browser automation, nor is it a software development tool. Picco engineer’s use Picco’s skills to make more skills. Picco has not yet reached the stage of being an autonomous agent, we don’t need to satisfy our customers. Currently, Picco lacks a conversational interface for customers—we are still relying on human coaches to translate customer requests into skills. The key is for Picco to deliver real value for customers, so we can achieve more agentic and AGI-like behavior over time.

What other key elements and design patterns do you think will emerge as application architectures evolve to incporate AI?

Excited to define a new AI First Architecture? Apply today.

CEO – Picco

CTO – Picco

Principal AI Enginner – Picco

AI Integration Engineer – Picco