Very senior software engineer & 1337 h@x0r ✨ Founder of @migusgroup Creator of https://teapawt.net/t.co/zFRgEMSMEo and https://teapawt.net/t.co/rEVGn8iSLi
Alexandria, VA, USA
Joined March 2009
- Tweets3.2K
- Following961
- Followers691
- Likes3.4K
Pinned Tweet
At long last, Lisa Harness is here! 🤖🎁
Lisa is a groundbreaking AI agent-harness:
- Hard budgets and enforced limits
- Dynamic model routing
- Lossless context (no compaction)
- Verifiable actions
Anthropic, OpenAI, et. al. are trying to create models you can trust. That's cool, but I'd rather run models without having to trust them. 🚫💀
There are a lot of other cool features, but TBH, the biggest W is its ability to keep many complex, multi-topic conversations straight and learn with you. 🧠
I've been working on it and using it for ~6 months. It's still beta quality; this is the first public release, but it's very usable and useful.
Blog post: migus.org/adam/introducing-l…
The project link is in my bio.
LMK what you think!
I was honored to be Cristina's first guest! It was great to talk about the trials and tribulations of being a technical founder!
LMK what resonates with you!
To celebrate reaching 3000 followers, and as promised, I'm bringing you today the first episode of my podcast, Founder Conversations!
My first guest is Adam Migus @amigus, an experienced software engineer, cybersecurity expert, and founder of CloudServers.
We discuss his journey from employee to independent consultant to SaaS founder, the challenges of turning a technically solid product into a profitable business, finding the right customers, marketing, pricing, and why building software is often the easy part.
We also talk about Lisa, his open-source AI agent harness, and the opportunities and challenges AI is creating for software developers and entrepreneurs.
It's an informal, unscripted conversation about what building a business actually looks like, beyond the usual startup success stories.
Thank you all for helping me reach this milestone, and I hope you enjoy our conversation!
Watch the full episode on YouTube:
youtu.be/njv7Yx2QJ2U
Adam Migus retweeted
To celebrate reaching 3000 followers, and as promised, I'm bringing you today the first episode of my podcast, Founder Conversations!
My first guest is Adam Migus @amigus, an experienced software engineer, cybersecurity expert, and founder of CloudServers.
We discuss his journey from employee to independent consultant to SaaS founder, the challenges of turning a technically solid product into a profitable business, finding the right customers, marketing, pricing, and why building software is often the easy part.
We also talk about Lisa, his open-source AI agent harness, and the opportunities and challenges AI is creating for software developers and entrepreneurs.
It's an informal, unscripted conversation about what building a business actually looks like, beyond the usual startup success stories.
Thank you all for helping me reach this milestone, and I hope you enjoy our conversation!
Watch the full episode on YouTube:
youtu.be/njv7Yx2QJ2U
Apparently we now publish {x}stack, where x = your initial and the stack is your agents and skills(?).
I wrote a whole harness on the premise that agents and skills don't work as a system of organization. So, you prolly won't see "astack" anytime soon.
Pitty. Such a cool name.
Lisa Harness has some users! 😆
Not many, so I'm not going to go all crazy with the RE process, but I did have to switch to full PR promotion- no more tag pushing. 😑
So v0.4.0 is a real release, and you should upgrade.
If you installed before `lisa update` came along, run the installer again or just `uv tool upgrade lisa-cli[prompt]` to get up to date.
If you're using it, bravo, you've been RT'ing that FM! 😆 HMU if you are interested in using it more and want me to think through your use case as I gear up with the tutorials and whatnot.
For all the power built into Lisa's model router, this is the fanciest I get with it nowadays:
version: 1
mode_models:
coding_design: kimi-k3:cloud
coding_execute: glm-5.3:cloud
coding_review: deepseek-v4.1-flash:cloud
default: beam:cloud
researcher: kimi-k3:cloud
shell: glm-5.3:cloud
router:
backend: lisa-router
rules:
- when: >-
route.operation_id in [
"consolidate_semantic",
"markdown_forge_document",
"synthesize_and_store_learning"
]
use: deepseek-v4-pro:cloud
Can you guess which provider I'm using based on the model IDs? 😜
Okay, GitHub. Now I'm genuinely curious.
Lisa Harness has had 1,377 clones from 179 unique cloners in the last 14 days.
Only 44 unique visitors to the repo.
One day: 442 clones from 35 unique cloners.
I know some of this is bots, scanners, indexing, etc.
But Lisa is an autonomous AI agent harness, so I have to wonder:
What exactly is cloning it?
Developers experimenting privately? AI coding environments? Agents examining another agent harness? Something else entirely?
I honestly don't know.
If you've cloned Lisa, I'd genuinely love to know what you're doing with it. 📥
And, you know... the ⭐ button is right there. 😄
When I built Lisa, I decided *not* to try making it a "personal assistant" because I thought Hermes and OpenClaw had it covered.
The thing is, everything that makes an AI agent a good PA is also a prerequisite for efficiency, reliability, and safety -- what I *did* focus on. So now, as it turns out, Lisa is an excellent assistant.
Lossless context and an immutable thread log are magic. Lisa keeps everything straight, no matter how many topics, never forgets, and always completes the assignment, whether it's coding, research, or server troubleshooting. 😎
Did I mention Lisa is FOSS under the MIT license?
github.com/Lisa-Harness/lisa
Lisa Harness can see the whole Linux machine but can't change anything on it. Knowing that, Lisa writes scripts for me and tells me what they do, why, and how to run them. 💾
Human-in-the-loop isn't a gate; it's a consequence of bash in bubblewrap on overlayfs. Lisa knows that and acts accordingly. 🧠
Why make agents trustworthy when you can just make agents that work without trust in the first place? 🤔
I built an agent harness host, Lisa on that premise and figured out the answer: you don't. 😲
But don't take my word for it. It's free open source so try it yourself! 🚀
I've got a lot of experience with sandboxes.
Here's my take on agent sandboxing, with a working example.
✅ Automatic
✅ Granular
✅ Minimal
✅ Transparent
Run your Linux box without risking it to a model.
Just follow the workbook to try it. 👇
github.com/Lisa-Harness/lisa…
A friend used Lisa to diagnose and fix a Docker issue. Yeah, I know, BFD, right? The thing is, Lisa had no write access. 🚫💀
We added the lisa-pack-bwrap. Lisa immediately diagnosed and fixed the bubblewrap issue on the box (an AppArmor conflict), gave us a script to fix it, and then, once we ran it, used a bubblewrapped bash session to resolve it by delivering a human-assisted plan, rationale, and bash scripts.
We could have also carved out access for Lisa to fix it directly, but we're building trust. 😉
I built Lisa so you *don't have to* trust the model. 🔒
github.com/Lisa-Harness/lisa
I posted about Lisa on my Drupal blog yesterday. Check it out: migus.org/adam/introducing-l…
It's slow because of all the traffic I'm getting... 😎
Lisa is really good at systems administration now.
Even using Bash in Bubblewrap, it can figure out what's wrong and help me fix it. We work through the sandbox to solve issues, and *I* take the consequential actions (i.e., frfr human-in-the-loop).
So, while I can turn off bubblewrap and give Lisa full write and sudo, I don't ever feel the need to, and *that* is the point. Lisa solves problems without needing dangerous levels of access.
I'm releasing it as FOSS this coming week, so don't be jealous; just stay tuned! 😄
What do y'all use to post video clips of the terminal/prompt?
I often post screen caps of Lisa's output, but I want to show her "in action," i.e., responding to a prompt...
Lisa, the FOSS agent harness I started developing in March, is almost done. It is modular, multifunctional and has a lot of cool features! 😊
The problem is I built it by myself and didn't build a community around it and I feel like I need to before I can release it.
Advice?
This was/is a defining feature of Lisa. The agents "own" their tools. Always knew it was powerful. Didn't realize it was catching on! 😊
Considering adding a recommendation to /retro to turn repeated tasks into custom tooling (i.e. a CLI + skill)
Talking to @poteto made me realise the power of agents delegating tasks to scripts they own and control
Means fewer tokens, fewer hallucinations