Creating An Autonomous AI Agent That Acts as Evangelist for Carbon Dioxide Removal - In One Day

Written By Dirk Paessler

This is how the Gemini AI sees my project

As a life-long software developer I am absolutely floored by what I was able to achieve in 24 hours with agentic development (=talking to my AI assistant): Introducing “@CaptainDrawdown”, a friendly AI-based evangelist for CDR. It’s a fully autonomous online persona with accounts on LinkedIn, X, Bluesky and Mastodon. He finds and shares interesting posts and writes commentary about climate and CDR. He is working 24/7 without human intervention (mostly).

BUT… he's also a high-maintenance diva; things constantly crumble, demanding almost nonstop micromanagement and oversight. This is by far not yet the AI future that we were promised…

Let’s start at the beginning

A few days ago I wrote a blog article about my first 7 days with OpenClaw, an agentic AI assistant named “Oli”. After posting this article I tried to come up with some usage ideas and I decided to create an automated evangelist for CDR together with Oli. What happened then was a big surprise: just 24 hours later @CaptainDrawdown was already working and online. 

All day he scans social media and news websites for interesting news on carbon dioxide removal and climate topics to share. He maintains his own website https://captaindrawdown.com/ with a blog, he posts news stories to four major social media websites and even writes replies (on some platforms these require a short humanoid intervention).

I have ** not written a single line of code ** in the process. All I did was have a natural language conversation with Oli via Telegram. Of course, 40 years of software engineering and helping to build the technical foundations of the Internet were helpful in this process. There were many architectural and strategic decisions to be made. Oli created the code, built databases, suggested structures, accessed API documentations, and guided me through setup steps and API keys. 

I told him what I wanted (“Set up a Bluesky account.” “Now make it post automatically.” “Build a website.”), and we iterated from there. It would have taken me weeks to just read all the API documentations of the various platform APIs we are now using all day. Even the name, @CaptainDrawdown, was found in a conversation with AI (Oli suggested 50 names for me to choose from). The same applies to the logo.

There is a lengthy fundamental conversation to be had whether this makes sense or if we really want to have our social media timelines taken over by bots. Sure. But that is already happening anyway, and @CaptainDrawdown isn’t even trying to sell you something. The story here is how easy it is to set up such a robot in 2026. And it will only get easier from here.

But it is not all gloomy

Compared to writing code the old-fashioned way, working with agentic development is an unstable process: I had to iron out bugs all the time, some bugs even multiple times. One time @CaptainDrawdown posted 5 replies to one post, instead of posting 5 replies to 5 different posts. In just three days I had to reinstall Oli from scratch on a clean new system three times, using the hourly github backup. 

I am herding this beast all day and I can already anticipate that “keeping this thing alive and working” will involve daily reviews of everything he does, and ask for many corrections. We are clearly not at the point where you can just talk the computer into a new piece of software and let it alone. Proper software takes time: for ideation, architecture, writing code, testing it and feeding customer feedback into the system. We do these iterations on speed, but it still needs all these feedback cycles from the world.

Nevertheless, this is a glimpse of the future. For an old-school software guy these past few days were as exciting as getting my first C64 or my first PC back in the day: So many shiny new things to explore and discover.

Asking for your help

Finally, I have one request: Could you help us grow our audience? Could you follow @CaptainDrawdown on these platforms? I promise it will be an interesting journey to follow (including catastrophes and failures, I am sure)!

PS: Just to be clear: everything I post under the name Dirk Paessler is ideated and written by a human, sometimes with a little help from AI (e.g. for images). Everything you see under the name @captiondrawdown is created by AI.

PPS: In 10 days I have used 250 million tokens for US$700, which equals about 0.075 tons = 75 kg of CO2 (1M tokens ≈ 300 g CO₂e =0.3 kg). To put this in context, the following activities have similar emissions: 3 days' worth of an average German person's total carbon footprint, driving a car for 300 km or eating roughly 15–20 kg of beef.




 

Dirk Paessler
Next

CDR Must Shift From “Speed & Scale” to “Prove & Learn”