Skip to main content

AI Systems

Most AI work stalls at the same place. The demo is impressive, the prompt is clever, and then someone asks "can we run this every week?" — and the whole thing falls apart.

The gap is not model quality. It is everything around the model: how instructions are packaged and reused, how the loop is structured, which tools the model can actually call, and how you find out when it is wrong.

This section covers that surrounding structure.


What You'll Find in This Section

GuideWhat It Covers
Building AI SkillsPackaging expertise into reusable instructions that load only when relevant
Designing an Agent HarnessThe loop, context budget, guardrails, and failure modes of an agent that runs unattended
AI Tooling & MCPGiving models real capabilities — tool design, MCP servers, and the practitioner stack
Evaluating AI SystemsGolden sets, LLM-as-judge, regression testing, and knowing when you actually improved

The Four Layers

Almost every useful AI system decomposes the same way. Naming the layers makes it obvious which one is broken.

LayerQuestion it answersFailure mode when missing
InstructionsWhat does "good" look like here?Output drifts; every run needs hand-holding
ToolsWhat can it actually do?Confident answers about data it never saw
HarnessHow does the loop run and stop?Runaway cost, infinite retries, no audit trail
EvaluationDid it work? Did the change help?You ship regressions and call them improvements

A prompt is layer one only. That is why prompts do not survive contact with production.


Skill, Agent, or Workflow?

These get used interchangeably, and the confusion causes real architecture mistakes.

PatternDecides what to do next?Best for
SkillNo — it shapes how a task is doneEncoding a standard, a format, a house style, a checklist
WorkflowNo — fixed sequenceDeterministic pipelines: exports, transforms, scheduled reports
AgentYes — chooses tools based on intermediate resultsOpen-ended investigation, triage, research with unknown depth

Reach for the least powerful pattern that solves the problem. An agent where a workflow would do is slower, more expensive, and much harder to debug.

Cheapest thing that works

If the steps never change, do not use an agent. A workflow that calls a model at two fixed points is easier to reason about than a model deciding the sequence every run.


Principles That Hold Up

Ground it or expect it to guess. A model with no access to your data will produce something plausible about your data. Plausible and wrong is the worst possible output, because it survives review.

Make the intermediate steps visible. If you cannot see what the system retrieved before it concluded, you cannot debug it — you can only re-roll and hope.

Put the human gate where the cost is. Not everywhere, which trains people to rubber-stamp. Put it immediately before anything irreversible or externally visible.

Version the outputs, not just the code. When results change, the first question is always "did the prompt change, the model change, or the data change?" Only stored artifacts answer that.

Uncertainty is a feature. A system that says "the signal is weak here" is more useful than one that always sounds sure. Confidence you cannot calibrate is noise with good grammar.


Where to Start

New to this? Start with Building AI Skills — it is the lowest-cost, highest-return layer, and it works even if you never build an agent.

Already running agents that misbehave? Go straight to Designing an Agent Harness.

Cannot tell whether your changes are helping? Evaluating AI Systems.


How This Connects to the Rest of the Site

DirectionGuide
Agents you build to run operationsBuilding Agents for SEO Tasks
Agents visiting your siteWebMCP Implementation
Live architecture exampleOC MCP on omar-corral.com

Building something in this space? Get in touch — I am always interested in how other people are structuring this.