Skip to content
LetsVibeAI

Vibe Coding/Module 7

Agents: harness, skills, tools, loops, goals and graphs

Module 7 — Agents: harness, skills, tools, loops, goals and graphs

An agent is an AI model that works in a loop: it reads instructions, decides what to do, uses tools, checks the result and repeats until the goal is met. You build one by organizing its parts into folders, then asking a coding agent to wire them together.

Lesson 7.1 — Chatbot vs agent

A chatbot answers one message at a time. An agent takes a goal ("research these 20 leads and draft emails"), plans the steps, calls tools (search, CRM, email), looks at what happened, and keeps going. Formula: Agent = Model + Instructions + Tools + Memory + Loop.

Lesson 7.2 — The agent vocabulary

Term Plain English
LLM The model doing the thinking (Claude, GPT, Gemini, Grok)
Model provider The company serving that model by API
API A doorway one program uses to ask another program for something
SDK A software development kit: ready-made code for using an API in your language (Vercel AI SDK, Anthropic SDK)
ADK An agent development kit: an SDK specifically for building agents (Google's ADK, Claude Agent SDK, OpenAI Agents SDK)
Agent harness Everything wrapped around the model — instructions, tools, memory, loop, guardrails — that turns it into a worker
Agent runtime Where and how the agent actually runs: the process that executes the loop, calls tools and streams results
Agent control panel The dashboard where humans watch, approve, pause and review agents
AI gateway One connection that routes requests to many model providers, with fallbacks, budgets and logging
MCP (Model Context Protocol) An open standard for plugging tools and data into any agent — "USB-C for AI"
Sandbox A locked-down, disposable computer where the agent can run real code safely, away from your files

Lesson 7.3 — The agent folder structure

Build an agent by giving each part its own folder. Start in Google Drive or your repo; type plain-language notes in each; ask your coding agent to fill in the rest and connect them.

/agent
  /instructions   what the agent does, its tone, rules for every situation
  /agents         the main agent's definition
  /sub-agents     delegated specialists (researcher, writer, reviewer)
  /skills         repeatable workflows, written as step-by-step recipes
  /tools          services it can use: email, CRM, search (often MCP servers)
  /functions      small single tasks reused across skills
  /knowledge      docs, user data, industry best practices
  /memory         how it remembers past sessions and users
  /harness        how all of the above is assembled
  /runtime        how it executes and produces the final artifact
  /gateway        which model providers and comms services it connects to
  /sandbox        where it runs code safely

Example: in /tools, make a doc that says just "email, CRM, calendar." Then prompt: "For each tool listed in /tools, create an MCP server or connector, with the auth it needs and a test."

Lesson 7.4 — Skills, processes and thinking systems

A skill is a repeatable process the agent can apply to any matching task — written once, reused forever. Example: a "write a landing page" skill with steps for research, outline, copy, review. Skills turn one-off prompts into a system. Most modern harnesses (Claude Code, Codex, Hermes, OpenClaw) load skills from a folder automatically.

Lesson 7.5 — Loops, goals and graphs

  • Loop engineering — the agent keeps working through a checklist until every item is checked, testing after each step. Your job is writing the checklist and the test for "done."
  • Goals — instead of steps, give a measurable goal ("all pages pass the accessibility check") and let the agent loop until it's met, with a cap on attempts and cost.
  • Graphs — a map of steps (nodes) and paths between them (edges), including loops back. Example: draft → review → (fail? back to draft) → publish. Graphs make complex agents predictable.
  • Human in the loop — pause for approval before anything irreversible: sending email, spending money, deleting data.

Lesson 7.6 — The Vercel agent stack

Piece Role Link
Chatbot template A complete, deployable chat app to start from chatbot.ai-sdk.dev
AI SDK TypeScript toolkit for calling models, tools and streaming ai-sdk.dev
AI Gateway One API for hundreds of models, with fallbacks and spend controls vercel.com/ai-gateway
AI Elements Pre-built chat UI components elements.ai-sdk.dev
Tools as packages Template for packaging reusable agent tools GitHub template
Workflow SDK Durable, multi-step workflows that survive restarts workflow-sdk.dev
Chat SDK Ship one agent to Slack, Teams, Discord and more chat-sdk.dev
Sandbox Isolated VMs to run agent-written code vercel.com/sandbox
eve Open-source agent framework: agents as folders of instructions, skills, tools and settings vercel.com/eve
Passport Single sign-on identity for internal apps and agents vercel.com/passport
Connect Short-lived OAuth tokens instead of stored secrets vercel.com/connect
Security Firewall, bot protection, DDoS vercel.com/security

Further reading: AI SDK templates, vercel.com/ai, and The Register's overview of eve and Passport.

Exercise: create the /agent folder for one agent in your app (e.g. a tutor or a lead researcher). Write 3–5 lines in each folder's notes file, then ask your coding agent to scaffold it with the AI SDK.

Check yourself:

  1. What five things make an agent?
  2. What's the difference between a skill and a tool?
  3. Why should irreversible actions have a human in the loop?

← Previous · Course home · Next →