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!
It took me forever to test "intent routing" with Lisa. So long that I formed a bad habit of not giving it a /mode at the start. A good model can default ->target_mode very quickly, but it still wastes time and tokens.
Also, "rage write?"
People do that?? 😂
1
91
This is another manifestation of "cognitive surrender," and its a trap.
You still write the code. You just do it through code generation with a model instead of typing it out "by hand."
Its still your responsibility.and *the key* your thinking.
Its only when you stop doing the thinking, thus surrendering to the tool, that makes you feel this way.
Don't fall for it.
Ever since AI started coding better than me three months ago, I’ve lost all hope for devs, and now I’m just bored writing code myself. My only real value is managing agents to get the best out of them. It feels like an engineering manager role, but I wanted to keep that hands-on ownership of solutions, which just isn't going to happen now that agents are writing better software than we ever could.
42
Lisa-Harness takes this approach.
The agent and harness are intertwined with the sandboxing controls that constrain them.
And, based on that direct experience, I'll say confidently that think he's spot on. ✅️
Sandboxes are the iFrames of agent execution.
Untrusted execution, loosely integrated, isolated, and separate.
Just like iFrames, agent sandboxes represent one of the coarsest forms of integration between the container (the harness or orchestrator) and the agent.
I believe that in order to achieve true AI composability, we’ll need to see a shift toward native agents that are deeply integrated with agent runtimes, with a much more fine-grained approach to agent and model execution, permissions, isolation, and security.
Sandboxes will remain important for executing untrusted operations, but they shouldn’t be the primary abstraction for composing AI applications.
1
88
I've been getting the benefits of flash models without derailing the work by starting with flash, then escalating to the full model after {x} iterations.
Obviously, it only works in a harness that can switch the model based on the iteration. You do it in the model manifest in Lisa-Harness, e.g.:
```yaml
mode_models:
default: glm-5.3-flash:cloud
coding_design: kimi-k3.1-pro:cloud
coding_execute: glm-5.3:cloud
router:
backend: lisa-router
rules:
- when: >-
route.mode_id in ["coding_review", "default"] &&
turn.iteration > 9
use: glm-5.3:cloud
```
24
I can't trust the model, so I built a harness that doesn't. And I am so glad I did; every day it saves my bacon. 🥓
1
26
I built Lisa-Harness with different constructs than agents and skills. It uses modes and operations instead.
I didn't want to because I knew it would be a headwind for adopters. But I had to because I needed more structure and boundaries than that structure offers, to create the unique features it has.
Its different but if you take a day to learn it, its worth the reward, I swear! 😊
19
"It's time to panic!"
K Einstein! F*cking AI doomers. 😑
It's time to panic about AI.
For years, I listened to the reasonable people. I nodded along when critics called AI “a con,” and when they waved ChatGPT away as “spicy autocomplete.” I laughed in 2024, when the internet was mocking ChatGPT because it couldn’t count the r’s in “strawberry.” When I met young researchers who had devoted their careers to the danger of rogue AI, I thought: what a waste of talent.
But at some point, I started to realize something was off. The machine kept learning things it was never supposed to learn.
OpenAI recently published more than 700 papers with the solutions to hundreds of open problems in mathematics. Hugo Duminil-Copin, winner of the Fields Medal (the highest prize in mathematics), wrote: “The shock is immense. I am paralyzed.”
None of this should have happened. So I went back and looked at what else the skeptics had predicted, and it dawned on me: they had been wrong about... pretty much everything. Next, I looked at the predictions of the crazy people, and realized they’ve been eerily prescient.
Why then, you may wonder, does the world still feel pretty normal? Because new technology usually spreads more slowly than its inventors expect (it took factories decades to reorganize around electricity). But the scary scenario is not about what happens in the wider economy; it’s about what happens inside a handful of companies.
I wrote down what changed my mind, and what we can do about it:
rutgerbregman.substack.com/p…
1
40
I started putting serious work into Lisa-Harness in March. I had a lot of ideas and requirements for it/her, but above all, I wanted an agent runtime I could trust.
I think it was probably May or June by the time I'd fallen victim to the overnight quota blowout- when my agent looped away all my credits with no solution at the end. But when I did, it only underscored the importance of the design decisions I'd already made. Like the token and tool-call limits.
Fast forward to now, and I'm still working out kinks and paving the way to a full software factory, but I *never* worry about blowing my quota or wasting tokens on unproductive loops! Those are solved problems in Lisa, and that feels amazing! 🥲
1
2
135
Learn to trust your agent/model? 🤨
Um, how about no? 😤
The AI engineering bubble is prone to groupthink, and this is the latest: The harness doesn't really matter. Just do xyz, and you can learn to trust your agent!
No "xyz" makes that statement correct because models are inherently untrustworthy! They lie and hallucinate. It's a fundamental truth of the technology. So harnesses matter more now than ever because agents cannot be reliable (or trustworthy) without them.
The harness creates the trust! ☝️
20
Do we really feel like the definitions of agent and harness are blurry and/or that the distinction doesn't matter? 🤨
I think the definitions are very clear and the distinction matters a lot! Who's with me? 🙋
1
1
44
I don't think AGI is real. And I definitely don't think models are (more) intelligent or somehow more capable than humans.
However, I do think AI is an extremely powerful tool that will completely remake software engineering and countless other STEM disciplines. It will make practitioners more effective and productive, and if, e.g., they're medical researchers, then it will accelerate medical advancements.
But let's please stop giving the tool credit. That'll only happen when the humans using it use it effectively towards those ends.
AI won't take anything from you unless you surrender to it, which, BTW, is the worst way to use AI.
11
Plan like you did with that sleezy PM you had at your last job. The one who just made sh*t up to make the plan *look* complete.
Because that's what a model is; a slick talker.
Best AI coding tip I've learned:
Don't let it code first.
Make it write the plan.
Fix the plan.
Then let it code.
Fewer bugs. Fewer rewrites.
18
Lisa's CLI is a bounded turn-by-turn process that you can loop (or not) instead of a service loop you mediate, to avoid these issues. 😏
25
Why? So that you can manage risk effectively.
Scary, a dude responsible for this much stuff doesn't get how cybersecurity risk management works. 😬
A friend recently prompted his own Linux ABI-compatible OS for his cloud applications. Why use off-the-shelf libc or device drivers or TCP stack, with all of the attendant bloat and attack surface? PromptOS is also much faster than stock Linux in production. I suspect that everyone will be doing this very soon.
1
43













