Who’s that Pokemon? 🦔

It’s agentic AI! 🤖

For the longest time, I thought AI agents were just fancy chatbots with extra steps or even an overrated transformer models. Then I started to follow the development my own and realized I was completely wrong.

Let’s say, agentic AI is basically Pokémon. Stay with me here.

Got the image from here

It’s starts with basic setup: I chose Charmender! 🐣

In Pokemon world, we all started with a starter. Mine was ChatGPT and its similar. As basic setup, you type something. It types back. Sometimes it hallucinates an entire research paper that does not exist (I have been there 😭). But you keep using it because, honestly, it is still pretty magical the first hundred times. You can generate so many texts within few hour, even with low credibility and long time to train.

Got the image from here

At first, ChatGPT was my Charmander. Small, basic, everyone’s starter. One time it tried to convince me that I published some articles already, but it kept saying that I never publish a thing. 🥲 I was so confused whether I wanted to keep investing my time with it or not. Then, I found Perplexity that is a tool that use Anthropic’s model Claude and make some chain of thoughts to find the evidences before stitching it together as an answer (which almost all AI platform doing it now). It was simple. Direct. No orchestration. No memory. Just prompt in, response out.

But starters are not meant to stay basic forever. At some point, you look at your Charmander and think, “Okay, cool, but can you actually do things?” Well, now it kinda evolves, even offers computer for me to work directly there. But it requires more subscription and to me, it feels a bit too much.

That is where the evolution begins.

It starts to evolve: Charmeleon 🔥

Got the image from here

So I went down a rabbit hole. My friend and I had just finished the Kaggle Google Agentic Bootcamp few months ago, where we built ZebraX, an agentic research tool using the Google Agent Development Kit. You can read more about that experience here. That project cracked something open for me. I saw what agents could do when they were not just answering questions, but actually executing tasks with customized chain of thoughts. Few months after, I kept going and keep curious with the same principles. Then, I tried CrewAI and LangChain. I wrote about the whole experience in “My weekend with new agents” if you want the full breakdown.

When testing CrewAI, I felt like I am managing a tiny startup inside my terminal. You define agents with roles and goals, assign tasks, and watch them pass work back and forth. It was kinda similar to my ZebraX project. But, the workflow was more straightforward and the foldering system was not needed. It was also a bit like herding cats, if the cats were Python scripts and the herding involved a lot of YAML 🙃 I also tested LangChain where it gave me more control. I built pipelines, defined chains, handled retries, and set up logging. It felt like engineering infrastructure. Powerful, flexible, but honestly a bit demanding (similar to Google ADK).

Both frameworks taught me something important:

Agents are not just prompts.

They are systems.

And systems need orchestration.

A 2026 study in npj Health Systems (Nature Portfolio) put numbers to this. Under clinical-scale workloads, multi-agent systems dropped modestly in accuracy where it drops from 90.6% at light load down to 65.3% with eighty tasks queued up. Single agents, by contrast, collapsed from 73.1% down to 16.6% over the same range (Klang et al., 2026). Same models. Same prompts. But the orchestrated ones kept going.

CrewAI and LangChain were my Charmeleon. More aggressive, harder to control, powerful but temperamental. You had to know what you were doing or you would just burn yourself.

But I was still thinking like a trainer with one Pokémon. Stronger, sure. But still one.

Then it becomes personalized: Charizard 🐲

Similar to the Pokemon’s story, I got frustrated and it’s kinda confusing when jumping from one to another project where I am keeps forgeting the class/ settings. So I did what any reasonable developer would do. 🫣

I built my own thing. 😂

Got the image from here

It is called SimAgents. I actually meant “Simple Agents,” but the name stuck and it was available on PyPI. So here we are, Charizard 🥲

The idea was dead simple: write a markdown file with your prompts, run it, done. No wiring. No boilerplate. No rebuilding the kitchen every time I wanted to cook.

SimAgents was my Charizard. Personal evolution. Streamlined, fast, and finally felt mine. For a while, I thought that was the final form.

I should have known better.

Current final form: Mega Charizard 🧬

Then I found OpenClaw and I liked it immediately. I spent one weekend building so many projects that my GitHub started looking alive again. The next few days, then I heard about Hermes. It has persistent memory and also popular in the developer community. It looked smarter on paper. So I did what any curious developer would do. I removed OpenClaw and installed Hermes.

That is when the disappointment started.

Removing OpenClaw wiped all my local settings as well (I should have made some backups 😭). Although, my projects were safe on GitHub and the containers are safe in my VM, but everything else was gone. I had to start from scratch.

Based on my experience when using Hermes, yes, it looked smarter. Yes, it could make skills by itself, which was genuinely impressive. But it was super slow. It overthought everything, probably just like me 😂 And the loop was endless. I would ask for one thing and it would spin for minutes, burning tokens like there was no tomorrow.

I felt like it was not for me. So I switched back to OpenClaw that same night. Just a few hours apart. Set it up again, this time with better security. I noticed a lot of Claude instances trying to poke around when I ran some security measures, so watch out for that.

Now I am back and I might stick with it for a while.

Here is what that actually my focus when rebuilding OpenClaw for me:

• My agents need to remember what I was working on lately
• They need tp schedule recurring tasks (like checking literature or monitoring experiments)
• They have to spawn subagents when a task needs specialized help
• When deliver results where I actually need them, it will send the stuffs to my telegram, email, whatever.

And yes, this is my Mega Charizard! 🧬

The image is coming from here

OpenClaw is my Mega Charizard. The final form. But only after you learn how to handle it. If you try to ride it before you are ready, you will get burned. And here is the lesson I keep relearning: both OpenClaw and Hermes are powerful. Neither is perfect. The tool is not the point. Learning to use it is.

The crazy part is that it keeps evolving. The skills system means I can add new capabilities without rebuilding anything. The cron jobs mean I can automate recurring work. The memory means my agents actually learn what I care about over time.

MemGPT, the paper that introduced OS-inspired virtual context management, put it best: stateful agents “remember, reflect, and evolve dynamically” through long-term interactions (Packer et al., 2024). That is the opposite of a chatbot that forgets everything the moment you close the tab.

I am not saying this is the final form. I have learned my lesson about calling anything “final” in AI. There will probably be another evolution next year. Or next month. Or even next week.

But right now?

This is the most capable setup I have used.

And the weirdest part?

It also feels the simplest.

Will I use it for work?🤖

Not yet, to be honest.

It’s more for hobby in my opinion. With work, it’s a bit tricky since all the data and setups need to be security measured and having those automations still a bit risky.

At work, it is better to have local LLMs or 3rd party which we have Non-disclosure Agreement (NDA) with. Then, I can keep sensitive data on-premise or the cloud that I know it’s safe. With persistent memory, I can maintain long-running analyses across sessions. With subagents, I can delegate different parts of a pipeline to specialized workers.

The privacy angle is not theoretical. A JMIR scoping review of 464 healthcare LLM studies found that 15.1% explicitly used local deployment to protect patient data, and 38.4% did not report their privacy measures at all (Xu et al., 2025). When you are dealing with GDPR and patient genomes, “trust the cloud” is not a strategy. It is a liability. Gartner also predicted that over 80% of enterprises would have deployed GenAI by 2026, up from under 5% in 2023 (Gartner, 2023). That wave includes local deployments, especially in regulated industries where a single data breach costs more than the entire LLM budget.

In my PhD, I spent weeks building analysis pipeline CellProfiler for detection, custom Python for classification and statistics and then figure generation. Half the time I could not tell which script version I was running or which parameter file was current. That fragmentation has a price. An analysis of over 250 developer teams found that tool switching alone consumes 20.5% of developer time, roughly 1.6 hours every single day (McAlpin, 2025). I feel that in my bones.

Well, now agentic memory changes the equation. The agent remembers last run’s parameters, flags which samples had artifacts, and adjusts the pipeline without being told twice.

The productivity case is also well-documented and studied. PwC surveyed 300 senior executives in May 2025 and found that 66% of companies adopting AI agents reported measurable productivity gains (PwC, 2025). When Stanford’s Digital Economy Lab asked workers what they most wanted from automation, 69.4% said the same thing: time back for high-value work (Stanford Digital Economy Lab, 2025). That is not about replacing humans. It is about letting humans do human things.

Another two cents from me 🪙

Do not chase the highest evolution just because it exists.

Got this from here

I have seen people jump straight into complex multi-agent systems before they even understood what a single agent can do. It is like trying to enter the Pokémon League with a Charizard you borrowed from a friend. Sure, it is strong. But you do not actually know how to use it.

Start simple. Use ChatGPT for what it is good at. Try a framework like CrewAI when you need coordination. Build your own tool when the existing ones frustrate you. And only then, consider something like OpenClaw when you genuinely need persistence and autonomy. The tools are not the goal. The work is the goal. The best Pokémon is the one that matches your gym badges, not the one with the highest base stats.

Also, there are things that you need to consider..

Testing things costs you time and money. In AI world, tokens are the currency, but buying them need money. I burned through credits learning to optimize, and when using ocal models, they are just limited. If you use it for OpenClaw/ Hermes, subagents fail silently sometimes. And I still cannot fully see what my subagents are doing, which bugs me. But the tradeoff works. I get more done with less management and continuity across sessions. A system that learns what I care about.

Imperfections are part of it. My Charmander missed moves, got confused, fainted. That is okay. You grow together and might end up with Mega Charizard! 😄

and my final thoughts…

The pattern never changes and it’s always shiny tools that get headlines. However, simple tools always get work done. So, beware with very new framework since people rush in (probably me as well 😂), burn credits (again me 😄), post on LinkedIn (also me 😂).

Three months later? Back to ChatGPT and a spreadsheet. Not because fancy tools are bad. Because simple was fine for the actual problem.

My advice:

Try everything, trust what works. Pick the tool that matches where you are, not where you wish you were.

For me, that is OpenClaw right now. Hands-off, persistent, and it lets me delegate without losing my sanity.

What about you?

What is your Charmander? And what are you evolving it into? 🚀

Until next time! 🤖

Reach me on LinkedIn — Immanuel Sanka if you’re interested in a chat! 😄 Let me know! About me? Check my first post here!

References

• CrewAI (2025) CrewAI framework. Available at: https://www.crewai.com/
• Gartner (2023) Gartner says more than 80% of enterprises will have used generative AI APIs or deployed generative AI-enabled applications by 2026. Available at: https://www.gartner.com/en/newsroom/press-releases/2023-10-11-gartner-says-more-than-80-percent-of-enterprises-will-have-used-generative-ai-apis-or-deployed-generative-ai-enabled-applications-by-2026
• Hermes (2026). Available at: https://hermes-agent.nousresearch.com/
• Klang, E., Omar, M., Raut, G., et al. (2026) ‘Orchestrated multi agents sustain accuracy under clinical-scale workloads compared to a single agent’, Nature npj Health Systems Research, 1, p. 77. Available at: https://www.nature.com/articles/s44401-026-00077-0
• LangChain (2025) LangChain framework. Available at: https://www.langchain.com/
• McAlpin, T. (2025) ‘Houston, we have a $20B productivity problem’, LinkedIn / Atlassian Research. Available at: https://www.linkedin.com/posts/troy-mcalpin-2a936b6_houston-we-have-a-20b-productivity-problem-activity-7364287837522808833-MAvN
• OpenClaw (2026) OpenClaw documentation. Available at: https://openclaw.dev/
• Packer, C., Fang, V., Patil, S.G., et al. (2024) ‘MemGPT: Towards LLMs as operating systems’, arXiv preprint arXiv:2310.08560. Available at: https://arxiv.org/abs/2310.08560
• PwC (2025) AI agent survey: the state of agentic AI in the enterprise. Available at: https://www.pwc.com/us/en/tech-effect/ai-analytics/ai-agent-survey.html
• Sanka, I. (2025) SimAgents. GitHub. Available at: https://github.com/im-sanka/simagents
• Sanka, I. (2025) ZebraX. GitHub. Available at: https://github.com/im-sanka/ZebraX
• Stanford University SALT Lab (2025) Future of work with AI agents. Available at: https://futureofwork.saltlab.stanford.edu/
• Xu, Z., et al. (2025) ‘Considerations for patient privacy of large language models in health care: scoping review’, Journal of Medical Internet Research, 27, p. e76571. Available at: https://www.jmir.org/2025/1/e76571