Rocky Lhotka

An Agentic Tale VP, Open Source Creator, Author, Speaker

An Agentic Tale

blog.lhotka.net/2026/03/10/A...

#ai #agentic #debugging #story

Agent Resources and Tools VP, Open Source Creator, Author, Speaker

An #ai #agent needs a good set of resources and tools to be useful. This blog post goes through the tools in the #rockbot framework and how they are used by agents.

blog.lhotka.net/2026/03/09/A...

GitHub - missingchildrenmn/KidsIdKit: This is a digital kids id kit app that allows parents to maintain a digital record of their kids' information. It is a simple app that allows parents to add their... This is a digital kids id kit app that allows parents to maintain a digital record of their kids' information. It is a simple app that allows parents to add their kids' information and prov...

Calling mobile devs! We're launching KidsIdKit, an open-source project for Missing Children MN. We need volunteers w/ expertise submitting apps to the App Store & Google Play.

Want to use your skills for good? Check out the repo & DM me! github.com/missingchild...

#TechForGood #iOSDev #AndroidDev

How RockBot Learns New Skills VP, Open Source Creator, Author, Speaker

#rockbot not only uses skills, but it develops and enhances skills over time so it is able to become more effective and helpful over time without manual intervention!

blog.lhotka.net/2026/03/06/R...

#ai #agent #dotnet #agentic

Stop Drawing Architecture Diagrams Manually: Meet the Open-Source AI Architecture Review Agents | Microsoft Community Hub Designing and documenting software architecture is often a battle against static diagrams that become outdated the moment they are drawn. The Architecture...

Stop Drawing Architecture Diagrams Manually - Meet the Open-Source AI Architecture Review Agents by Shivam Goyal techcommunity.microsoft.com/t5/educator-...

So you know someone who wants to get into fullstack development with #dotnet? Do then a favour and send them this link

www.coursera.org/professional...

The RockBot Band VP, Open Source Creator, Author, Speaker

To be useful, #rockbot needs other agents and #mcp servers, because it follows the principle of least privilege. Each has permission to do just what it does. #agentic #ai in action.

blog.lhotka.net/2026/03/03/T...

A few of these aren't out of the oven yet, but they will be this year.

Meanwhile, be sure to ask your FLGS to help you fill in any gaps in your collection.

This interview is one of my most honest and upfront. I also wax poetic a few times bc I’m old. Anyway, thank you Shingai Njeri Kagunda. Great questions!

One common challenge with #agentic systems is agent/tool discovery. To solve this, I built a registry service for #a2a, #mcp, #acp, and message-based #a2a, so virtually any agent/tool can list itself for discovery within a system.

github.com/marimerllc/a...

Last night #rockbot noticed that I have some conference submissions (calls for speakers) open, so it created talk titles/abstracts for me to submit.

Very thoughtful 😎

It is a real challenge for sure! Lots of prompt tweaks and in-code guard rails, and it still isn't entirely reliable.

Just like a lot of humans I know 😏

Why am I creating #rockbot? Because I want to have a personal/professional #ai bot that I can trust - that has clear security boundaries and which could be connected to my business services in addition to my personal and work emails and calendars.

github.com/MarimerLLC/r...

I love how #claudecode says something will "take a day of focused work" and then plans and implements the entire feature in 30 minutes.

Software development has become so much fun again!

You where you have a child give you instructions on how to make a sandwich and you literally do what they say? It is hilarious!

Building an #ai bot isn't all that different, but instead of hilarious, it is frustrating.

The rockbot framework tracks when a memory was last touched so it can expire memories that don't get brought into context. Also, when it dreams, memories are often consolidated or refined to be more useful.

How RockBot Remembers Many AI models have no persistent memory. Every conversation starts fresh. They don’t remember what you told them yesterday, last week, or five minutes ago in a different chat window. For a casual ass...

A discussion of the #ai #agent memory system used by #rockbot to remember conversations, activity, collaboration, and long-term memory.

blog.lhotka.net/2026/02/24/A...

It was some temporary outage sort of thing - this 401 error, because now (without warning) I have access to the 1M context window again - whew!

This goes to show what we lost when twitter died though - no central place to get immediate answers to this sort of issue.

Lesson learned: use at least #claude haiku if your agent uses tools, cheaper models hallucinate way too much.

#rockbot is starting to shape up pretty well today; good tool use, subagents, #a2a agents - truly productive #ai.

github.com/marimerllc/r...

Did #anthropic just change access to the 1M context window models? I was busily making cool stuff, and then all of a sudden the 1M models (#sonnet / #opus) started returning 401, so I had to switch to their smaller versions.

GitHub - MarimerLLC/rockbot: Autonomous personal agent designed to be cloud-native Autonomous personal agent designed to be cloud-native - MarimerLLC/rockbot

#rockbot now has a _ton_ of new features after the weekend! Supports #a2a protocol, subagents, background tasks, 3 levels of memory, and has major identity updates.

I gave up trying to use #deepseek and have been using #haiku, and that helps - worth the extra cost imo.

github.com/marimerllc/r...

#Claude Desktop does a lot of nice things Microsoft Copilot doesn't do, so I use it a lot.

#ClaudeCode has a massive ecosystem that #githubcopilot doesn't match - skills, plugins, etc.

I do use the Claude models via Copilot all the time, but the actual Claude tools are also better imo.

It is really sad that #claude with #cowork now consumes 10% or more of the system CPU to do absolutely nothing. I have to remember to close Claude Desktop after every interaction or my PC is so sluggish it is frustrating.

GitHub - MarimerLLC/rockbot: Autonomous agent designed to be cloud-native Autonomous agent designed to be cloud-native. Contribute to MarimerLLC/rockbot development by creating an account on GitHub.

Like a human needs to dream, so does an autonomous agent. #rockbot is designed with a dream subsystem so it can organize all sorts of memories, skills, and other information to make its life better and to be more productive when it "wakes up".

github.com/MarimerLLC/r...

I'm trying to make #rockbot able to compensate for "cheap" models like deepseek terminus, but it is really hard! I'm not convinced that it isn't better to just pay for a good model like Claude sonnet or opus.

It makes me think of the bad old days when we had to deal with random browser quirks and people building web sites had farms of PCs to test different browsers and configurations.

Yuck!

Defeating LLM hallucinations is a real challenge, to say the least. They are different per-model, so it is necessary to use a per-model provider to deal with them.

Geez, dealing with time and time zones is almost as hard as dealing with precision of double values and rounding!

I discovered yesterday, that if you are running as root it won't allow you to --dangerously-ignore-permissions. So there's still _some_ safety :)

Just had my 10 day ICD checkup and the device is working great, which explains why I've been feeling so much better. Modern medicine, science, and technology are pretty amazing!