Comans Services
Practical guide
Getting Started with AI Agents
Comans Services • Practical guide

Getting Started with AI Agents

Real tools, real workflows, stuff you can use today. No tech background needed. If you can write a clear email, you can do this.

Built for people who want practical AI, not hype.
Use it tonight, not someday.
Join our next workshop

Want the live version of this with demos, Q&A, and practical examples? Join the next Comans AI workshop.

Join workshop waitlist

Start here: pick your path

Getting started with Claude or ChatGPT

Option 1: Claude (our recommendation)

  1. Go to claude.ai
  2. Sign up with your email
  3. Start chatting. The free tier is enough to learn.

Why we recommend Claude: Better at long documents, more natural writing style, excellent reasoning.

Option 2: ChatGPT

  1. Go to chat.openai.com
  2. Sign up with your email or Google account
  3. Start chatting

Why you might pick ChatGPT: Bigger ecosystem of plugins, image generation built in, more guides online.

Option 3: Google Gemini

  1. Go to gemini.google.com
  2. Sign in with your Google account
  3. Start chatting

Why you might pick Gemini: Integrates directly with Gmail, Google Docs, and Calendar.

Your first conversation

Summarise
I need to summarise this article for my team. Here’s the text: [paste article]. Give me a 3-sentence summary and 3 key takeaways.
Draft
I need to reply to a client who’s asking about project timelines. The project is 2 weeks behind but we have a plan to catch up. Write a professional but honest email.
Brainstorm
I run a small accounting firm in Melbourne. Give me 5 ideas for attracting new clients that don’t involve cold calling.
Explain
Explain how superannuation works like you’re talking to a 25-year-old at a barbecue. Keep it under 200 words.

Prompt patterns that actually work

Most people type vague questions and get vague answers. These four patterns fix that.

Pattern 1: Give it a role

Weak prompt

Help me write a business proposal.

Better prompt

You are a business development manager at a mid-size IT company. I need a proposal for a potential client who runs a chain of medical clinics. They need help with cybersecurity and cloud migration. Write a one-page proposal that focuses on reducing their risk.

Pattern 2: Show, don’t tell

Weak prompt

Write me a professional email.

Better prompt

Here’s an email I wrote last week that got a great response: [paste your email]. Now write a similar email for a different client, [describe the situation]. Match the tone and structure.

Pattern 3: Think step by step

Weak prompt

Should I hire a new staff member or outsource?

Better prompt

I’m deciding whether to hire a full-time staff member or outsource to a contractor. Here’s my situation: [details]. Think through this step by step, considering cost, flexibility, quality, and timeline. Then give me your recommendation with reasoning.

Pattern 4: Add constraints

Weak prompt

Tell me about cloud security.

Better prompt

Explain the top 3 cloud security risks for a small business, in plain English, in under 150 words. No jargon.

Combining all four

You are a nutritionist (role). I want a weekly meal plan for someone doing intermittent fasting, eating one meal a day around 6pm (context). Here’s an example of a meal I enjoyed last week: grilled chicken, sweet potato, and a big salad with feta (example). Give me 5 different meals, each under 600 calories, as a simple list with ingredients and estimated prep time (constraints).

Common mistakes

  • Asking vague questions and expecting precise answers.
  • Pasting sensitive data into free AI tools.
  • Treating AI output like truth instead of checking it.

Subscriptions: what’s worth paying for

What model should I actually use?

  • Claude Sonnet: fast and smart for daily use.
  • Claude Opus: better when the work is more complex and you want deeper thinking.
  • ChatGPT-4o: a solid all-rounder if you want one general model.
OptionCostBest forPrivacy
Claude Free$0Trying it out, light useMay train on data unless you opt out
ChatGPT Free$0Trying it out, light useData may train models
Claude Pro~$33.99 AUD/moDaily use, long documents, better modelsCan opt out of training
ChatGPT Plus$20 USD (~$33 AUD)Daily use, Custom GPTs, image generationCan opt out of training
Gemini Advanced$32.99 AUD/moGoogle ecosystem integrationGoogle’s privacy policies apply
Ollama (local)$0Maximum privacy, no internet neededData never leaves your machine
API access~$8–$75 AUD/moDevelopers, agents, automationData not used for training

Our recommendation: start free, move to Claude Pro at about $34/month if you use it daily, and use Ollama when privacy matters most.

How to install Ollama (local AI, zero cost)

Mac

  1. Go to ollama.com
  2. Download the Mac app
  3. Open it. It runs in the background.
  4. Open Terminal and type:
ollama run llama3.2

Windows

  1. Go to ollama.com
  2. Download the Windows installer
  3. Run the installer
  4. Open PowerShell and type:
ollama run llama3.2

Popular models: llama3.2, mistral, codellama, nomic-embed-text. As a rough guide, llama3.2 is about a 2GB download and runs fine on most modern laptops.

Privacy and safety

Warning

If it’s data about a person, a client, or a business, and it’s identifiable, don’t put it into a free AI tool. Use a paid plan with training turned off, or run it locally.

Never paste

  • Client personal data
  • Financial records
  • Medical information
  • Passwords or API keys
  • Legal documents with sensitive details
  • Anything you would panic about in a data breach

How to use AI safely

Level 1 - Basic: turn off training, use anonymised examples, treat the AI like a public whiteboard.

Level 2 - Better: use paid plans, keep training off, and review all output.

Level 3 - Best: use Ollama locally or API access for sensitive work.

Setting up your first agent

Option 1: Claude Code

Claude Code is Anthropic’s terminal-based AI agent. It can read files, write code, run commands, and take actions on your computer.

Requires Node.js 22 or later. That’s the most common install issue people hit.

npm install -g @anthropic-ai/claude-code

Option 2: OpenClaw

OpenClaw is an open-source AI agent platform that turns your AI into a 24/7 assistant.

curl -fsSL https://openclaw.ai/install.sh | bash
openclaw onboard

Option 3: Open Jarvis

Open Jarvis-style self-hosted assistants are the voice-first home play: private, local, and customisable. Best fit if you want a Siri/Alexa alternative that runs on your own hardware instead of sending everything to the cloud.

  1. Start with the project docs or GitHub install guide for your chosen Jarvis implementation.
  2. Use a small always-on device like a Raspberry Pi or spare mini PC.
  3. Connect a microphone and speaker.
  4. Run a local model with Ollama if privacy matters most.
  5. Begin with simple jobs: reminders, questions, smart-home actions, and calendar prompts.

This path is more DIY than Claude Code or OpenClaw, but it is the best fit for people who want a private home assistant rather than a desktop work agent.

OpenJarvis official repository

Agentic services vs human-based services

Advanced section

This section is more advanced. If you're just getting started, skip ahead to Daily Workflow and come back to this later.

There’s a shift happening that most people haven’t noticed yet. The tools you use every day, like email, CRM, project management, and customer support, are splitting into two versions. One is designed for humans clicking buttons. The other is designed for AI agents calling APIs.

Understanding the difference matters because it changes how you set up your business for the next five years.

What’s the difference?

A human-based service has a graphical interface. You log in, click around, type things, drag and drop. Gmail, Salesforce, Trello, and Xero are all built for human eyes and human hands.

An agentic service has an API or protocol interface. Instead of a person clicking buttons, an AI agent sends structured requests. It reads data, takes actions, and reports back without needing a screen.

The same platform often offers both. Gmail has a web interface (human) and a Gmail API (agentic). Salesforce has its dashboard (human) and Agentforce (agentic). The difference is who’s driving.

Why agents need installed services, not just web apps

When you use ChatGPT in a browser, it can’t touch your files, your email, or your calendar. It’s trapped behind glass.

For an AI agent to actually do useful work, it needs:

  • API access: structured ways to read and write data in your tools
  • Authentication: permission to act on your behalf (OAuth, API keys)
  • Local tool access: ability to run commands, read files, execute scripts on your machine
  • Protocols like MCP: Model Context Protocol lets agents connect to your tools through a standard interface

This is why installed agents like OpenClaw and Claude Code are more powerful than browser-based chatbots. They can actually reach your systems.

The big platforms going agentic

PlatformAgentic productWhat it doesWho it's for
CloudflareAgents SDK + Workers AIBuild and deploy AI agents on the edge with state, scheduling, tool use, and MCP support.Developers building custom agents
Cloudflare + Coinbasex402 FoundationAI agents that can autonomously make and receive payments using HTTP 402.The future of agent commerce
MicrosoftCopilot Studio + Agent 365Build custom agents inside Microsoft 365 across Teams, Outlook, and SharePoint.Businesses on Microsoft 365
SalesforceAgentforcePre-built and custom AI agents inside Salesforce CRM and service workflows.Businesses on Salesforce
GoogleAgentspace + Vertex AIEnterprise agent platform with Google data, search, and Gemini models.Businesses on Google Cloud
AWSBedrock AgentCoreBuild agents on Amazon infrastructure with models, knowledge bases, and action groups.Businesses on AWS
ServiceNowAI AgentsAgents built into ITSM, HR, and service workflows to auto-resolve tickets and requests.Enterprise IT and HR teams

Cloudflare’s approach is worth understanding

  • Code Mode: API surfaces described compactly so agents can act with far less token overhead.
  • MCP Server Portals: compose multiple services behind one gateway with unified auth.
  • Durable state: each agent instance can keep its own persistent memory and state.
  • Pay for CPU time, not wall time: changes the economics of agents that wait, wake, and act.

Agentic services: pros

  • Agents can work 24/7 without human involvement
  • Faster execution: API calls in milliseconds instead of human clicks in minutes
  • Consistent: agents don't forget steps or have bad days
  • Scalable: one agent can handle many tasks at once
  • Composable: agents can chain services together automatically

Cons

  • Require technical setup
  • Larger security surface
  • Debugging is harder
  • Costs can be unpredictable
  • Not everything has an API yet

Human-based services: pros

  • Familiar and easy to understand
  • Visual confirmation of what you did
  • Easier to catch mistakes in real time
  • No technical setup required
  • Good for one-off tasks

Cons

  • Doesn't scale
  • Error-prone
  • Human time is expensive
  • Not available 24/7 without shifts
  • Repetitive work burns out good people

The practical takeaway

  1. Keep human-based services for judgement calls, relationships, creative work, and anything where context and empathy matter.
  2. Move to agentic services for repetitive processing, monitoring, triage, first drafts, scheduling, and predictable workflows.
  3. Use MCP and APIs to connect agents to the tools you already use instead of replacing everything.
  4. Start small. Connect one tool, automate one workflow, measure the time saved, then expand.

The businesses that win won’t be the ones that go all-in on agents or reject them entirely. They’ll be the ones that figure out the right split between human judgement and agent execution.

Making AI part of your daily workflow

Morning routine (5 minutes)

  • Summarise unread emails
  • List your calendar and flag conflicts
  • Give yourself a 3-sentence industry briefing

During the workday

  • Draft first versions of emails and proposals
  • Prep for meetings
  • Turn rough notes into minutes
  • Think through stuck problems step by step

Weekly review

  • Summarise the week
  • Suggest next priorities
  • Review documents for gaps

The bigger picture

What AI is good at

  • Processing large amounts of text quickly
  • Drafting first versions
  • Summarising and organising information
  • Brainstorming and explaining
  • Code and language tasks

What AI is not good at (yet)

  • Judgement that depends on experience and relationships
  • Understanding your business context without being told
  • Guaranteeing factual accuracy
  • Replacing empathy, creativity, and leadership
  • Operating unsupervised on high-stakes tasks

The honest truth: AI replaces tasks, not people. Keep the human work. Automate the repetitive bits.

What to try tonight

  1. Sign up for Claude or ChatGPT
  2. Pick one boring task from today
  3. Ask AI to help
  4. Compare the result to the manual version
  5. Notice the time difference

The people who get the most from this won’t be the ones who wait until they feel ready. They’ll be the ones who try it, get a few clumsy answers, and learn fast.

Resources