From 6846cf1920844b11110fb91edeae06de65c0bb96 Mon Sep 17 00:00:00 2001 From: Eddie Soehnel Date: Thu, 16 Apr 2026 17:05:41 -0600 Subject: [PATCH] Update README.md Added notes about how to classify execution vs services layer vs harness/orchestration layer and where LLMS might fit --- README.md | 36 +++++++++++++++++++++++++++++------- 1 file changed, 29 insertions(+), 7 deletions(-) diff --git a/README.md b/README.md index 763d0e3..5f6ba5f 100644 --- a/README.md +++ b/README.md @@ -16,7 +16,7 @@ This framework guides the full lifecycle: how to define a product, how to build I organize the layers according to the 8 core functional areas of a business that I created, [as defined here](https://eddiesoehnel.com/startup-roadmap/). -### Tools: +### Process Boundaries: * If humans need to think through it → Google Docs to allow for shared working * If AI needs to read it to build something → Markdown in a repo @@ -29,14 +29,27 @@ The most valuable stack is: 3. data modeling 4. AI orchestration -### 3-layer system +### 4-layer system -1. Knowledge layer (specs/) → what the system is supposed to do and how to do it. The truth via highly structured executable knowledge that AI can understand. Structure leads to intelligence. Smarter AI models help, but if poorly structured, then failure. Well structure knowledger layers can succeed even with average models. -2. Harness layer (harness/) → the program that runs the AI. It does four things: runs the model in a loop, reads (specs/, tasks/, data-ops/), writes files (srv/), manages context (specs/), and enforces safety. That's it. It is "thin", very tightly scoped to a specific task. +1. Knowledge layer (specs/) (skills) → what the system is supposed to do and how to do it. The truth via highly structured executable knowledge that AI can understand. Structure leads to intelligence. Smarter AI models help, but if poorly structured, then failure. Well structure knowledger layers can succeed even with average models. +2. Harness layer (harness/) → the program that runs the AI. It does four things: runs the model in a loop, reads (specs/, tasks/, data-ops/), writes files (srv/), manages context (specs/), and enforces safety. That's it. It is "thin", very tightly scoped to a specific task. It sits between the srv/ and the AI. It is what activates the AI to operate in the srv/. Deterministic code still runs crons, acts based on conditions, executes if / then. That still exists, but now we have added AI as intelligence, so it uses the harness to operate. 2. Product layer (tasks/ and srv/) → takes the knowledge layer and executes it: develop code, develop the product, produce the product. 3. Data layer (data/) → what the system observes, stores, and learns from based on production. The data validates reality, that leads to iterate above to improve. -Intelligence is in specs/ (skills docs). Execution is in the srv/, and execution is deterministic. Deterministic is code calls, code execution, cron jobs. It just does, but does not think. +### Intelligence Vs Code +Intelligence is in specs/ (skills docs). + +Execution is in the srv/, and execution is deterministic. Deterministic is code calls, code execution, cron jobs. It just does, but does not think. If the task is exact and repetable, use deterministic code. + +If the task requires judgement or interpretation, then it is a skill. It use to be that we had to force everything into code - logic that was fuzzy, required interpretation and still terrible at intelligence. Now, we can collapse that cognitive logic that was in a lot of code into skills. LLMs introduce massive intelligence into the process that never existed. So, now, we have code that executes and scales, but now we can layer on intelligence that we never had before. + +Instead of trying to encode complex reasoning in deterministic code (if/then logic), we move that reasoning into skills that guide the LLM. We keep all execution, control flow, and precision tasks in code, and only offload judgment and interpretation to the model. + +Code scales execution and guarantees correctness. + +Skills scale intelligence and adapt automatically as models improve. When the model improves, every skills improves automatically without changing the code. That is how we can get 10x-100x improvement from the same system, with an improved AI model. + + ### Structure For Git Repo @@ -59,11 +72,20 @@ Intelligence is in specs/ (skills docs). Execution is in the srv/, and execution *The executable structure layer, organized by main function and in order* -- agents.md \ Functionas more like a router, telling AI where to find stuff. General rules. Set up clear path where AI agents orchestrate other AI agents. When typx X task appears, load document y first. You are pointing to documents that get loaded at the right time. You do not put all of this into one document. Keep context tight. -- skill.md \ teaches model how to do something; supplies the process for doing something; +- agents.md \ Functionas more like a router, telling AI where to find stuff. General rules. Set up clear path where AI agents orchestrate other AI agents. When typx X task appears, load document y first. You are pointing to documents that get loaded at the right time. You do not put all of this into one document. Keep context tight. Think of this as an organization chart. It defines who handles what, how tasks get routed, what happens when something does not match. It is escalation logic: when one path fails, where does it go next? +- skill.md \ teaches model how to do something; supplies the process for doing something. Think of skills as employees. Each one has a capability, Some are specialists, some are generalists. Some only run on cron. - decisions.md \ why we did what we did, why we chose A over B, etc. Architecture Decision Records (ADR). provide context, decision, risks, consequences. Document using the [decision layer and biases layer](https://docs.google.com/document/d/1VjtP-jPn-wOpE8z5QSVOdk-pPxneyuDDM6z_a9OIpYY/edit?tab=t.0) - soul or personas.md \ who the agent is. As it may run sub-agents, this document defines the personas for those sub-agents. - there could be many versions of these spec files in each spec folder area. +- consider filing rules - what goes where when it gets recorded. + +#### 3-Layer Classification System for "Agents" + +1. Execution Units is something that does work: py script, cron job, API call, function +2. Service Layer is a reusable unit that performs a defined function: +3. An agent is a system that decides what to do next. It is the harness that calls a LLM model to help apply judgement and reasoning about what to do next. +- It does not have to call an LLM - it could be a rules-based decision tree or a predefine flow. We could call this an automation orchestrator. +- And if judgement and reasoning are required, we can call this an LLM orchestrator, because now it calls for an LLM. ###### 1_strategy_s