Show Me Love

There's a lot of AI-related content coming up, and with it comes a ton of new terms that can be confusing. I put together a list of common ones and what they mean to help you understand the technology around them.

---

Model = trained to generate content

• supplies intelligence
• used for coding tasks
• drafting emails tasks

Tokens = units of text

• chunks of text that are often one word, part of a word or a punctuation mark that the model processes
• a prompt “hello world” can be equivalent to 2 tokens (depends on the model’s vocabulary)

Credits = billing unit

• does not exist in the model layer
• product layer of the provider/ vendor like Kiro
• e.g., the Kiro Power plan has 10,000 credits

Context = content that is loaded to the model

• conversations
• instructions
• document
• code
• image

Context Window = maximum amount of tokens the model can work on

• model’s working memory
• a 200K context window means that the model can process 200K tokens (input and output combined)

MCP (Model Context Protocol) = a standard to connect to vendor tools, features

Tool(s) = function(s)

• can be invoked via MCP
• can be done via Skill

Skill(s) = folder with procedures in a markdown file

• a standard to give agents capabilities
• discovered by agents on startup
• activated when task match
• can be executed with templates (sample files) and scripts (bash, python)

Agent(s) = Model + Harness

Harness = Agent - Model

• anything that is not a model
  • system prompts
  • tools
  • state
    • memory across time
    • session
  • loop
    • to keep going until the goal is reached
  • guardrails
    • allowed to do
    • boundaries
• supplies agency
• e.g. OpenCode, Pi, Claude Code, Kiro

Harness Engineering

• building a system around the model
• Troubleshooting:
  • model issue
    • hallucination, drift, confused
    • model reasoned wrong
  • harness issue
    • did something not supposed to
    • wrong setup
• Important notes:
  • engineer around the model
  • model is constant
  • model is swappable
  • harness to control
  • as the model gets more capable, the harness should get simpler
• Analogy
  • horse → model
  • straps → harness
  • working horse → agent
  • driver → human

Agentic Workflow

• deterministic outcomes
• predefined processes with AI embedded
• AI assists with the structure

Autonomous Agent

• agent decides the path
• goal driven
• plans and executes

---

Don't let me show cruelty Though I may make mistakes