There is a version of your week where the expenses are done before you think to do them. Where you arrive at every meeting already briefed. Where the research you need is sitting in your inbox before you ask for it. Where the admin that currently lives at the bottom of your to-do list, guilt-tripping you every time you scroll past it, simply gets handled.
That version of your week is available right now. Most people do not have it because they are using the wrong mental model.
They are treating AI and agents like software. Open it, type something, get output, close it. The same loop as every other tool.
That is not what a truly agentic experience is. Most people do not have agents. They have a chatbot with a slightly better interface.
Think of them as employees
The shift that changes everything is this: stop thinking about AI agents as software you operate and start thinking about them as people you hire.
Not metaphorically. Structurally.
When you hire someone, you do not hand them a laptop and walk away. You onboard them. You give them context about how things work, what matters, who to talk to, what decisions they can make alone and where they check with you first. You give them access to the tools they need. You define their scope. You invest time up front so they can be genuinely useful from then on.
An agent needs exactly the same things.
Give them context and they become calibrated to your world. Give them tools and they can act in it. Give them guardrails and they become reliable rather than unpredictable. Skip any of those three things and you get the generic, slightly-off output that makes most people conclude the technology is not ready.
The technology is ready. The onboarding is not.
Think of your agent as a brilliant new hire. High IQ. Fast learner. Genuinely motivated. Available at any hour. But completely new to your world and entirely dependent on you to tell them how it works. A first week with no induction produces a confused employee. A first week with proper context produces someone who is immediately useful and gets better every day.
The same is true here.
What a properly set-up agent actually does
Here is what people who have done this properly are experiencing.
Their agent reads their email each morning and surfaces the three things that actually need attention today, not a list of everything, the ones that matter, in the order they matter. It drafts responses in their voice, not a generic professional voice, their voice, the way they actually write.
It prepares them for every meeting. Not just the time and the dial-in. The relevant context: what was discussed last time, what was agreed, what changed since, what they promised and have not delivered. They arrive informed. Every time.
It handles the tasks that never get done. The expenses sitting in the downloads folder get processed. The subscription that should have been cancelled six months ago gets flagged. The follow-up email that got buried comes back up at the right moment. The research task that kept being pushed actually gets done, overnight, while they were sleeping.
It synthesises health data. Recovery scores, sleep quality, training load, cross-referenced with the calendar. One message each morning that tells them what they need to know about the day ahead, drawn from four different apps they used to have to check separately.
It manages communication channels. Inbound enquiries get handled. Scheduling conversations get managed. Requests get routed. The people who need a response get one, in a voice that sounds like the person it is representing, because it has been briefed to do exactly that.
One person cleared ten thousand emails on day one with a single instruction via WhatsApp. Another rebuilt their entire website from their phone while watching television, without opening a laptop. A professor described passing his expenses folder to his agent and having the whole submission done in minutes, a task he had been procrastinating for weeks because the university format was tedious and the instructions made no sense. He does not think about it anymore.
That last one is the one that matters most. Not the dramatic stories. The quiet removal of the tasks that drain energy without producing anything worth the energy they drain.
The channels you already use
One more thing worth understanding before we get into how to build this.
Agents do not live in a new app. They do not require a new dashboard or a new interface or a new habit.
They live in WhatsApp. Telegram. Signal. Slack. iMessage. Discord. The channels you already have open all day. You message your agent the same way you message a person, and it goes and handles what you asked.
This matters because every tool that has ever required you to go somewhere new eventually got abandoned. The behaviour change was too large. The habit never formed. An agent that lives in WhatsApp does not require a behaviour change. You are already there.
How to actually set one up
This is where most content about agents gets vague. So let us be specific.
Before we get into the components, two things worth saying plainly.
First, this is not exclusively for developers. The tools available now, OpenClaw, Claude Cowork, and others, are built for people who understand what they want rather than people who understand how to code. The setup requires thought, not technical skill. The hardest part is writing an honest briefing document about yourself and how you work. Most people find that uncomfortable because they have never had to articulate it. That is also why the agents built by people who do it well are so much more effective than those built by people who rush it.
Second, giving an agent access to your email, your calendar, your files, and your communication channels is a meaningful act of trust. Worth acknowledging rather than glossing over. The answer is not to minimise the access but to be deliberate about it. Connect what the agent genuinely needs. Define clearly what it can and cannot do with that access. Review its actions until you are confident in its judgment. The same way you would with a new employee handling sensitive information.
A properly set up agent has four things: a soul, a memory, tools, and boundaries.
The soul document is a file, usually called soul.md, that describes who you are and how you work. Not a CV. A briefing document for a new hire. Your priorities, your working style, your communication preferences, your pet hates, what good looks like in your world. Two pages written honestly. This single document is the difference between an agent that sounds like you and one that sounds like everyone else.
The memory file carries the live context. Current projects. Key relationships. Recent decisions. Open threads. Things the agent needs to know to be useful today rather than just in general. Some agents update this file themselves as they work. When that happens, the feeling is different. It starts to feel less like a tool and more like someone who has been paying attention.
The tools are what allow the agent to act rather than just advise. Email access. Calendar. Documents. Your CRM if you have one. The apps that hold the information relevant to the work it is doing. Without tool connections an agent can think and suggest. With them it can do. The distinction is everything.
The guardrails define where the agent operates independently and where it checks with you first. What can it send without review? What does it always show you before acting on? What does it never do without explicit permission? This is not about distrust. It is about design. Clear operating limits make an agent more useful, not less. A new hire who knows exactly when to escalate is more valuable than one who either does nothing independently or charges ahead without judgment.
Then there is the model: the brain you plug into the agent. There is a wide range available, each with different reasoning capabilities, different speeds, and different strengths. Claude, GPT, Gemini, open-source models like Llama and Qwen. None is universally right for every task.
What is becoming standard practice is routing your agent through something like OpenRouter, a platform that directs different tasks to different models depending on what each requires. Heavy reasoning goes to a more capable model. Routine drafting or data retrieval goes to something faster and lighter. You set the logic. The result is an agent that is both more capable and more efficient than one locked to a single model regardless of what it is doing. The quality of the model matters. But it is the layer above the model, the soul, the memory, the tools, the guardrails, that determines whether the agent is genuinely useful or just impressive in a demo.
Skills are modular capabilities you add as you need them. Web research. File management. Calendar intelligence. Health data integration. Social media drafting. Each one extends what the agent can reach for. You start with a few and add more as you understand what you actually need.
What we have built
Neo is a personal operational brain. Every morning, before the day starts, there is a message. WHOOP recovery data read and interpreted. Calendar reviewed. Three things that need attention. News filtered to what actually matters for the work currently in progress. Open threads from yesterday that need closing. Twenty minutes of context-gathering compressed to thirty seconds of reading.
Vicky is a Chief of Staff operating via WhatsApp Business. She handles inbound enquiries, manages scheduling conversations, routes requests, and maintains continuity across projects and relationships. She sounds like someone who has been doing this for years. Because the context she carries represents exactly that.
For a professional services firm: an agent that handles meeting preparation, competitive briefings, proposal first drafts, and project status across every active engagement without anyone having to compile it.
For an events organisation: an RFP agent that reads incoming briefs, maps them against the organisation’s history and capability, and produces a structured first draft in the organisation’s voice. Every proposal it has ever written, accessible. It knows what worked.
For a fast-growing company: a recruitment agent that benchmarks applications against role briefs, identifies the strongest candidates with explanation, and drafts the initial outreach. The decision stays with the humans. The noise is removed before it reaches them.
We are currently working with a general manager of a business to replace the entire administrative layer of his week with agents that operate independently. Scheduling. Briefing. Reporting. Follow-up. Work that previously required either his time or a dedicated hire.
None of these required months of development. They required clear thinking about what the agent was for, who it was working on behalf of, and what it needed to do its job properly.
Why now
The capability is here. The tools are accessible. What is missing, for most people and most organisations, is the framing.
Once you stop thinking about agents as software and start thinking about them as employees, the rest becomes straightforward. You know how to hire someone. You know how to brief them. You know how to give them the right tools and define their scope and trust them once they have earned it.
Apply the same thinking here.
Everyone should have a personal agent. At minimum, a personal assistant or a chief of staff that handles the work that currently fills your time but does not require your particular judgment or presence. Not because it is impressive. Because experiencing it properly, with real context and real tools, is the fastest way to understand where this is going.
It is going everywhere. Every organisation. Every function. Every role.
Getting into it now, building the intuition for what works and why, is one of the most useful things you can do this year. Both for what you can produce and for understanding how to navigate what comes next.
The agents are available. The question is only whether you have given them what they need.
If you want to explore what this looks like in practice for you or your organisation, drop a comment below or send me a message directly. This is exactly the kind of conversation I enjoy.
If this landed, subscribe. The next piece goes into the architecture in more detail: memory files, skill design, tool connections, and the specific decisions that separate agents that are genuinely useful from ones that sound impressive and produce nothing.
Craig Hepburn is an AI strategist and Perplexity Fellow who builds advanced agentic systems. He spent years leading digital transformation at Art Basel and UEFA. Now he works on the harder question: not whether organisations adopt AI, but how they govern it when they do.
Hi Craig, your posts in this area have been striking a chord. I’d love to connect and have a chat.
I saw you speak at Leaders Meet Innovation and found it fascinating! I would love to know how to go about building an AI agent for personal use. I have no idea where to even begin.