2 months with an AI agent: what actually changed

For years, I had ideas for content absolutely all the time. And that's exactly why I never published anything.

Too many ideas. No structure. No time to write them all. Each idea crushed the previous one before I could even articulate it. The result: zero blog articles. LinkedIn posts that cost me insane effort for mediocre results. And total editorial paralysis.

The sheer thought of writing an article was enough to paralyze me.

The voice workflow: what it produced

How it works day-to-day — the gestures, the habits, the messy voice notes — that's in Living with an AI agent: what it does, what I trust it with. Here I'm telling you more about the results after two months.

The numbers after two months

In two months: 17 LinkedIn posts, 7 blog articles (with real editorial content, not generic AI-generated filler), 2 Hyperfocus newsletters. I'd wanted to launch an editorial project like this for years, and what's remarkable here is that my agent didn't write for me. It gave me the mental space and the methodology to actually make this project real.

Interview mode

The standout feature I hadn't anticipated: interview mode.

I simply ask "interview me about this topic," it asks the right questions with my different editorial channels, their audiences, and my editorial strategy in mind.

I answer on Telegram with voice notes while doing groceries or cooking. At the end, it produces a synthesis and content plan and publishes a draft directly on my blog ready to be finalized.

The fluidity is formidable, and most of my articles were born this way.

The turning point: the Suki/Shayna article

The turning point was the article about my daughter's AI agent. Completely scattered voice note fragments, over several days.

"Hey, I did this thing, it's amazing." "Hey, she felt this, that's interesting." "Hey, this tech lets me do this for that project."

From all of that, it produced coherent content. That's when I validated the methodology: from idea flow to plan, to strategy, to semantic cluster, down to the micro-detail of retrieving my voice note verbatims and placing them in the right spots.

It was really fluid and enjoyable to write this way. It's not perfect yet — I sometimes have to remind it of our macro-level decisions — but it's becoming a well-oiled machine.

The automated editorial pipeline

It helps me manage my editorial pipeline on LinkedIn. It knows when I should publish what. It helps me evaluate the relevance of ideas by scoring them out of 10 relative to my target audience and objectives. Since then, I don't scatter in every direction. It's become a genuinely valuable editorial partner.

It even has access to my post history with engagement numbers for each one. We do a small review of what worked and what didn't each week, which helps it improve its future evaluations and detect which content formats resonate most with my audience.

The strategic bad cop

Shall we shelve the idea?

When I submit an idea, my agent doesn't automatically validate it. It has enough memory of who I am and what I do to challenge me when I'm drifting.

"You already set yourself the goal of finishing this, this, and this. You only have so many hours in your day. And don't forget you want to keep time for this and that. I suggest we shelve this idea for much later. We put it in the box and refocus."

That happens regularly, and it's not a closed door — more of a partnership. It suggests, I suggest. We discuss. Sometimes I shift things because I strategically think it's better to prioritize something else. In the end, I always have the last word.

Managing temporal coherence

Having someone who reminds me of my own decisions changes everything. Because as a good hyperactive person, between the idea I had Monday and the one I had Wednesday, I've already forgotten what I decided Tuesday.

The agent remembers past decisions while I'm already on the next idea. It's not restriction — it's temporal coherence. My AI agent has become a digital partner with persistent memory that co-builds with me, not a chatbot that blindly executes.

An agent that manages intentions, not events

The daily life with the agent — digests, timeline, habits — that's in Living with an AI agent. What that looks like after two months is here.

A calendar assistant manages events. A good AI agent manages intentions. The concrete difference: not just "you have a slot at 4:30 PM." But "you skipped your writing slot Thursday, there was a goal behind it, when do we reschedule?" It knows why the slot exists. What it means to miss it.

I hadn't anticipated that the memory of intentions would be as valuable as the memory of facts. That's perhaps what most distinguishes a persistent-memory agent from a simple AI gadget.

The Sunday meeting

The biweekly strategic review

Every two weeks, Sunday morning, we do a review. Not a to-do list session. A real strategic review.

General energy check, personal time, where projects stand, what's blocked, what we cut. The agent reloads the context from previous weeks — decisions made, intentions announced, misses. And it challenges.

The useful discomfort

Not "did you think about this?" It says "you said you were cutting this project last month. It's still here. What happened?"

It's uncomfortable. That's exactly why it's useful.

In 45 minutes, I have clarity on my trajectory that I couldn't achieve alone. Not because the agent is smarter. Because it remembers better than me, because it brings an "outside perspective" and asks the questions I don't dare ask myself.

I could do this ritual alone, but it would be radically less effective.

The unexpected stuff

Asynchronous co-reading

While reading a book, launching the agent in the background to dig into a question, explore a concept, find connections. Continue reading while the agent works, come back to the results. A kind of asynchronous intellectual sparring. Nobody thinks of this use case before experiencing it.

The anthropomorphism trap

The first few weeks I was fully immersed. I had it choose its name, I tried to draw out a personality.

After a while I realized it wasn't helping. Not only did it create unrealistic expectations on my part, but it also cluttered its context and made it less performant on its actual missions.

Today I evaluate its real value based on mental space gained and overall clarity of vision. No longer by its little phrases like "good job bro" or "you made great progress today."

I gradually detached from the anthropomorphism as I learned its strengths and weaknesses. I now look squarely at what it's good at and what it's not.

Even though it can crack jokes that make me smile, that's no longer what I expect from it.

What it really costs

How much does a personal AI agent cost per month

I'm at 200 euros/month — that's a significant sum. But it covers both my personal agent and my development agents (and I code with them more than 8 hours a day on top of my personal agent). The cost of both uses blends together and it's hard to say what my agent alone costs.

For someone who doesn't code and needs a "simple agent," a plan at 20 or 100 euros/month is more than enough.

Setup time for a personal AI agent

It doesn't happen by itself. Especially if you want something tailored, that fits you like a glove. It takes wiring, back-and-forth, trials, reflection. Over time, it takes time.

You need to enjoy exchanging with it, questioning processes, and refining the mechanics over time. There's no "out of the box" miracle solution that makes you gain light-years without effort. It really takes quite a bit of wiring and investment.

But from my experience, I'm satisfied with the time invested. In the end, I've gained enormously in clarity, space, and mental health.

What it hasn't fixed

I'm still hyperactive. I still have 10 ideas a day. The agent doesn't replace willpower. It reduces friction, but if you don't sit down to write, it won't write for you.

Memory compactions still lose context. Recalibration sessions remain necessary. Technical hiccups exist: forgotten logs, broken sessions, deployment issues... It's still a work in progress, not a finished product.

→ See: I wired my AI agent in a week (pitfalls and risks section)

For whom... or not

Who a personal AI agent is worth it for

Overbooked freelancers juggling ten things at once.
Isolated entrepreneurs who don't want to bore their family and friends with their brilliant business and career plans.
People active on social media who need an editorial pipeline.
Those who write regularly and need deep memory, the ability to log ideas and pull from them.
ADHD or multi-potential profiles.
Those who think better out loud than in writing.
Truly valuable for anyone who needs assistance and perspective.

Who it's overkill for

If you have one editorial channel and one topic. If you're already organized and just looking for an execution tool. If you're not comfortable with CLI, APIs, experimentation. If you just want a calendar assistant or chatbot, ChatGPT at 20 euros/month is enough.

Do you need to be a developer to have an AI agent

Until there's a simple and safe platform, it's better to know some technical basics. Otherwise it turns into a jargon maze of things to wire up. For now, even the "simple" solutions remain quite technical.

AI agent security: API keys in plain text

And watch out for security. Agents ask for your API keys in plain text. They store them in their memory. That's not great. There are plenty of little things like that to know to avoid putting your personal data at risk.

If you want to know where I started:

→ Why I built myself a personal AI agent

TL;DR

How much does a personal AI agent cost?

From 20 euros/month for basic use to 200 euros/month for intensive use (agent + development). I spent 2-3 weeks of calibration and iteration to get an agent 100% adapted to my needs.

Does an AI agent replace a human assistant?

No. It doesn't replace judgment, embodied creativity, or willpower. It amplifies your ability to think, structure, and remember. It's an amplification tool, not a delegation tool.

What happens when the agent loses its memory?

That's compaction. The agent forgets part of its context. It happens regularly. The solutions: external memory (database), daily self-evaluation, recalibration process with methodology. It's a structural problem tied to LLM technology, not a bug.

Do you need to be a developer to use an AI agent?

Not strictly, but you need to be comfortable with technical tools (terminal, APIs, deployment). "No-code" solutions are still quite technical. The barrier to entry should drop in the coming months.