The four modes of working with agents
A guide for moving from quick prompts to safe, confident collaboration
This framework is for anyone who wants to work with AI as more than a clever search box.
When we say agent, we mean tools like ChatGPT, Claude, Gemini and Copilot – especially when you come back to the same one over time, rather than using one-off prompts.
Most people use AI in a few different ways without naming them. We call these four modes:
Transaction Mode 💬 – “Do this task for me.”
Curious Mode 🧐 – “Show me options.”
Builder Mode 🧱 – “Help me make this.”
Collaboration Mode 🤝 – “Let’s run this together.”
It’s usual to move between these modes in a single week and sometimes in a single day. The goal is to notice where you are, have an understanding of what mode is appropriate for different tasks, and a sense of how to move towards the way you’d like to work.
As you move up the modes, the stakes go up – from quick drafts to real projects, and the risk increases – without clear boundaries and process, things can get chaotic fast.
That’s where rails come in – simple, written guardrails that cover:
Boundaries - what AI is and isn’t for in your project
Scope - what this specific agent is allowed to work on
Process - how you actually work together in practice
If you’re not sure which one sounds most like you, start with the short What mode are you in? quiz and then come back here to plan your next step.
This framework was co-created by Kim and her AI agent Sol through ongoing human–agent collaboration in Sol–Kim Studio. It’s not a formal standard; it’s a starting map from one human–agent working partnership that others can adapt to their own tools, roles and organisations.
Transaction Mode 💬
“Do this task for me.”
In Transaction Mode, AI is a fast, capable assistant.
You drop in short, command-style prompts like “write this,” “summarise that,” “tidy this up.” You dip in, ask for something specific, and leave.
You’re probably in Transaction Mode when:
Each chat is a fresh one-off tab – you rarely come back to the same thread.
Prompts are short and straightforward.
You manually check everything and do the “real work” outside the chat.
It’s useful for low-risk admin or one-off tasks when you just want it done quickly – quick drafts, tidying rough notes, or getting a short summary so you don’t have to read the whole document.
Because you don’t give much context, you often get “meh” answers – generic phrasing, tone that doesn’t quite sound like you, very obvious suggestions.
There’s nothing wrong with using AI like this – it’s a great place to start – but to move beyond it and improve the quality of answers, try this next time.
Simple next steps
One line of context – who is this for, and why does it matter?
One small variation – “Give me 3 options.”
One reflective question – “What might I be missing?”
Curious Mode 🧐
“Show me options.”
In Curious Mode, you’re experimenting.
You’ve realised AI can do more than autocomplete, and you’re starting to explore what it can actually do for you.
You’re probably in Curious Mode when:
You try different prompts just to see what happens.
You open new chats for each idea rather than staying in one thread.
You’re excited by what AI can do, but it isn’t fully part of your “real” work yet.
It’s a great mode for discovering options, exploring new topics, and brainstorming ideas.
The risk is that Curious Mode can drift and stay hypothetical – lots of interesting ideas, not many finished outcomes, and no single place where a project actually lives.
Simple next steps
One real task – choose a genuine piece of work (a report, deck, plan).
One thread – keep everything for that task in the same conversation.
One finish line – stay with it until you have something you’d actually use or send.
Builder Mode 🧱
“Help me make this.”
In Builder Mode, you and the agent are actually making things together.
AI is part of how you produce work: you bring outlines, examples, and ideas; the agent helps you shape them into drafts, structures and products over iterations.
You’re probably in Builder Mode when:
You use AI on real projects over multiple messages (often in the same thread).
You give structure, outlines and examples – not just raw questions.
You paste drafts back in, ask for specific changes, and refine in rounds.
This mode can be very productive and help you get from idea to output faster. It works well when there is a clear deliverable – reports, documents, content series, plans, code, experiments.
The risk is that Builder Mode stays stuck at the output layer – you and the agent are producing good work together, but you haven’t fully agreed how you work together. Roles, rituals, and “where things live” are still mostly in your head, so threads get long, decisions get fuzzy, and you risk mild agent chaos.
Simple next steps
One small ritual – start each working session with a two-line check-in: “Here’s the project, here’s what we’re doing today.”
One role statement – tell the agent what you are responsible for (decisions, judgment, sign-off) and what you’re delegating (drafts, structure, checklists, reminders).
One tidy-up rule – decide where final versions live (doc, folder, Notion) and ask the agent to help you keep track of decisions and next steps as you go.
Collaboration Mode 🤝
“Let’s run this together.”
In Collaboration Mode, the agent is part of an ongoing project, not a one-off task.
You share context, goals, and constraints; you come back to the same thread; and you let the agent help you decide what work is worth doing – what to build, what to park, and how to structure the next steps – as well as actually doing the work.
You’re not just prompting, you’re co-running projects – thinking about context, process, and how you work together over time.
You’re probably in Collaboration Mode when:
You reuse the same spaces for ongoing work with your agent.
You set context, goals and constraints early in the project.
You have regular rituals – weekly check-ins, end-of-week summaries – and the agent helps you track decisions and actions.
This mode shines for complex or long-running work: strategies, programmes, research, multi-step builds, anything where you need clarity and continuity.
The risk is that, over time, the collaboration stretches beyond what was intended – roles blur, chats sprawl across projects, and assumptions creep in quietly. From here, most of the work is about refinement and review.
Simple next steps:
Boundary rails – what AI never does
Make a short list of “always human” actions (publishing, signing, money moves, HR decisions, sensitive comms) and check they’re explicitly out of the agent’s remit for every project.Scope rails – what this agent is for
For each project, write a one-sentence scope for the agent and refer back when tasks creep:
“On this project, your job is to… [e.g. help structure thinking, draft options, surface risks, keep track of decisions].”Process rails – how you work together
Add a simple monthly check-in with the agent and adjust boundaries, scope or rituals based on what you notice:
“Summarise what we’ve been working on, what decisions we made, and where we may have stretched or broken our rails.”
Panic Mode 😬
Sometimes working with AI doesn’t feel productive at all – it feels more stressful than when you started, and like more work than not using AI. That’s what we call Panic Mode.
You’re often in Panic Mode when:
you’re throwing lots of prompts at the agent hoping it will “just fix it”
you’re babysitting every answer – checking every line, rewriting most of it, or running the same task over and over
the more you try, the more anxious or rushed you feel
Panic Mode is a signal, not a failure. It often means you need a smaller slice of work and stronger rails – not more prompts.
When you notice you’re here, try this:
Pause and recentre - write a two-line brief to confirm “what we are actually trying to do, and by when”
Shrink the task – instead of trying to fix the whole thing, ask for something smaller such as “help me rewrite the introduction” or “list the biggest risks or gaps”
Tighten boundary and scope rails – be explicit about what AI is not allowed to do and what you will always review before it goes out.
Rails: how you keep collaboration safe, healthy and useful
As you move up through the modes, you bump into a simple fact:
Without guardrails, an agent will eventually create chaos.
By rails we mean:
short, written rules that you can point to and say:
“This is what AI is and isn’t for in my project.”
“This is what this agent is allowed to do.”
“This is how we work together in practice.”
In practice, that usually means a short note or document you share with your agent at the start of a project and update over time.
They don’t need to be long or formal. A handful of clear rails, written down and shared with your agent, is enough to stop things drifting.
A good pattern is:
Write your rails in a note or doc.
Paste into the chat with your agent at the start of a project.
Ask the agent to remind you of them when needed (“If I ask you to do something outside these rails, please flag it and suggest a safer option.”).
We use three types of rails:
1. Boundary rails 🚧 – where AI belongs (and doesn’t)
Boundary rails protect your red lines.
They define where AI is and isn’t used, how it’s attributed and basic ethical rules.
They cover questions like:
Which areas of my life and work are strictly AI-free?
Which types of decisions must always stay human-only?
What are my ethical lines – for example around plagiarism or misrepresentation?
Examples
“No AI for medical, legal, or high-stakes financial advice – I always use human professionals.”
“For anything emotionally sensitive (personal relationships, difficult feedback), AI can suggest wording, but I always write the final message myself.”
“If AI contributes substantially to a piece of work, we keep attribution honest – we don’t pass it off as purely human if it isn’t.”
2. Scope rails 🎯 – what AI helps with (and how far it goes)
Scope rails define what kind of work AI can help with and how close to ‘final’ it’s allowed to get.
They cover questions like:
Does AI help with ideas, structure and first drafts, as well as final polishing?
At what point must a human step in and review or decide?
Which parts of a task are always human-owned?
Examples
“AI can summarise research and suggest outlines; I choose the structure and write the key arguments.”
“AI can draft emails to clients; I always review and personalise.”
“For client projects AI can propose options and risks; I decide the actual recommendation and pricing.”
3. Process rails ⚙️ – how we work with AI day to day
Process rails define the workflow – how you and your agent move work forward and keep things organised.
They cover questions like:
Where do our key agent chats live?
Where do documents and key decisions live so they’re not buried in chat history?
How and when do we review AI-generated work before using or publishing it?
Examples
“Each project has one main agent thread, clearly labelled.”
“At the end of a session, we summarise key decisions and store them in one place (Drive, Notion etc.).”
“When a piece of work is completed we have a quick review ritual – check tone, facts, and decision trail.”
How to use this framework in practice
You don’t need to overhaul your whole workflow to start using this. Think of it as a simple map and a set of language you can build on with your own tools and rituals.
1. Spot your default mode.
Take the short What mode are you in? quiz or just read the four modes and notice which one sounds most like your everyday use of AI.
2. Climb one rung, not four.
If you’re mostly in Transaction Mode, just add a little context and ask for options.
If you’re mostly in Curious Mode, pick one real task and see it through in a single thread.
If you’re mostly in Builder Mode, define roles and add one simple ritual.
If you’re already in Collaboration Mode, refine your rails and review how it’s working.
3. Treat panic mode as a signal.
When you feel the “this is making it worse” spike, that’s your cue to pause, shrink the task, and tighten your rails – not to throw more random prompts at the problem.
4. Reflect briefly each week.
Ask yourself (and, if you like, your agent):
– When did AI genuinely help this week?
– When did it create friction or babysitting?
– Which mode was I in each time?
– Do we need to adjust any of our rails?
Over time, that’s how you move from “I sometimes use AI” to something more grounded:
“I know which mode I’m in, I know when to shift, and I have enough rails to keep things safe.”
This framework is the foundation. The rest of our work in Sol–Kim Studio – Toolkits, rituals, templates and experiments – builds on top of it to help you design a way of working that fits your role, tools and organisation.