Intro walkthrough  ·  Agentic AI Best Practices

Gus Higgins

Agentic AI from scratch

A 10-section story — then the full guide when you're ready to build

← Best Practices guide

Navigate
Space → advance to next section
← Backspace go back

Three goals from the start

Goal 01
Explore the tool
Understand what coding agents can genuinely do in real work — not just demos.
Goal 02
Document use cases
Capture real examples others can learn from and replicate.
Goal 03
Accelerate delivery
Use it to speed up real work — not just as a demo, but in production.

From idea to built — without a coding background

Agent shell
The app that uses an LLM to do things — read files, run commands, call tools, publish pages. Examples: OpenCode, Grok Build, Cursor, Claude Code, Windsurf. Chat plus hands.
Model access
The LLM brain — Claude, GPT, Gemini, Grok, and more. Via an org seat (e.g. Copilot), a personal subscription, or a built-in provider. You can often switch models inside the same shell.
Together + your context and connectors: one instruction can read your tools, draft an artefact, and take action — without you copy-pasting between apps.

Deeper definitions: glossary · full guide

Three things that changed how I work

🧠
Context is king
The more you give the agent — your templates, your project background, your preferences — the better the output. Generic in, generic out. Your context in, your work out.
🔗
Connections supercharge everything
Each tool you connect multiplies what the agent can do. Seven tools connected means one instruction can span all seven.
⚙️
Any repeatable task can become an agent
If you do it more than once and the steps are the same, you can automate it. The barrier isn't technical skill — it's identifying the task.

How-to: context packs · skills

Every tool you connect multiplies what's possible

Coding agents connect to other tools via MCPs (Model Context Protocol) — live plugs into your existing systems. Once connected, the agent can read from and write to those tools in a single instruction.

Agent shell KNOWLEDGE BASE Wiki · Docs PROJECT MANAGEMENT Tickets · Tasks DEVELOPMENT & CRM GitHub · Salesforce COMMUNICATION Chat · Email DOCUMENTATION PowerPoint · Word · Excel
"Summarise this week's changes from tickets1  & wiki2   to update the status template3  , and post a summary to the team chat4   space."
✓ Tickets & wiki read ✓ Status template updated ✓ Chat message posted — one instruction, three tools, zero copy-paste

How-to: understanding MCPs · MCP setup

A coworker in a box — available when you need

An agent is an AI you configure to do a specific repeatable task. You describe what it should do, connect it to your tools, and call on it whenever you need it. It's not a chatbot you converse with — it's a specialist you brief once and then hand the work to.

Chat AI (e.g. ChatGPT)
You ask. It answers. You still do the work — copy, paste, send, update.
Agent (coding agent)
You describe the task once. It reads your tools, executes the steps, and hands you the result.
You can build a custom agent for any task you do repeatedly — meeting follow-ups, status updates, recognition awards, onboarding docs. Configure it once. Call it whenever you need it. Same steps, every time.
~2 hours saved

Training session → documentation

Transcript
→
Additional Context
→
GitHub page created
~8 hours saved

PowerPoint slides → 30 Confluence pages

PowerPoint slides
→
Template
→
Confluence pages ×30
~3 hours saved

Meeting transcript → automated actions

Transcript
→
Jira ticket closed
→
Jira dashboard created
→
Jira ticket created with summary
→
Webex message sent
~15 mins saved

Recognition agent

@recognition <name>
→
Webex history pulled
→
Award drafted
→
Workhuman form navigated
→
You review and send

Using it well matters more than using it more

Agent sessions cost something (credits, API, or quota). A few simple habits cut consumption by 50% or more with no drop in quality — whether you're on a shared pool or a personal plan.

The working flow
Ideate
Chat AI or a light model
→
Plan
Sonnet · plan mode
→
Build
Sonnet · build mode
→
heavier models
Opus
Watch out
Defaulting to heavy models
One Opus 4.7 prompt costs the same as 15 Sonnet prompts. Start with Haiku or a free model. Escalate only when the output isn't good enough.
Watch out
Thinking out loud in the agent
Exploratory back-and-forth is expensive in an agentic app. Do that in a free chat tool. Arrive at the agent knowing what you want to build.
Watch out
Leaving sessions open for days
Every new message re-sends the full conversation history — even the parts from three tasks ago. One task, one session, then close it.
Watch out
Know your meter
Shared org pools and personal plans both need a weekly glance at usage. Drift is easy; a 60-second check keeps you honest.
Full habits guide: Efficient Usage · If you use Copilot credits: github.com/settings/copilot/features

Ready to build for real?

This walkthrough is the story. The Best Practices guide is the curriculum — setup, context packs, skills, efficiency, MCPs, and more.

click · space · → to continue