Adam Migus

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!

Sandboxes are built into Lisa's execution layer so there's nothing to spin up. Just saying.

I follow Josh because he's always explaining why all the things I did in Lisa are the best thing to do! 😊 I'm an OS guy so, naturally when I did subagents, I created 'fork' and 'delegate' with the former doing exactly what he points out as best case. 💪

ℹ️ This common misunderstanding in harness engineering is critical to avoid if you want to reason about sandboxing, agent security and the like. 👇

This is the world I built Lisa Harness for. I'm glad I saw it coming. 😎

This, literally overnight, became my biggest worry. 😨

This is what I'm worried about. Call it cognitive surrender or whatever you want. Deferring to models is not a good thing to do!

Lisa is now a software factory! 😯 But not mine; I don't have a reason to turn my process, which is nothing special anyway, into a factory, so I decided to use a better base: HDAE from @gmkurtzer. I figured if he could build an OS kernel with it, it's a solid methodology to bring into Lisa. Of course, his work is published as a collection of agents and skills and, if you're following along, you know Lisa doesn't do agent skills. So, the work is more of a port than an integration. Either way, the HDAE "pack" lets Lisa follow the process, but per Lisa, the harness (not the model) enforces the gates and verifies the results. That's why you'll see far fewer agent prompts than in Gregory's Codex setup. I'm willing to bet that, with a skilled operator like him, it could build the same thing with way fewer tokens. But I don't have the token budget to prove that out myself. 🫤 But hey, if you do, please go for it and HMU if you need any help; I'd love to be proven right or wrong! Oh, and if you think this doesn't count as a software factory for some reason, tell me why! I'm still not entirely sure of the definition myself... 😬 The pack is available in the 'lisa-pack-hdae' repo in the Lisa-Harness GitHub. Just add it to your session along with 'lisa-pack-bwrap ' and 'lisa-pack-coding' to get started. 🚀

📢 I'm going to release a FOSS software factory built on Lisa this afternoon! The twist is that its not based on my factory design. 🤔 Stay tuned if you're into harness engineering and agent factories! 👀

Wow, that's like 4 quote layers of wrong. Impressive.

Lisa’s turn loop 👇 I'm going to assume anyone reading this knows what a turn loop is. If not, hey, follow me on LinkedIn! 😆 🔁 The loop always has an active mode that determines its instructions, available capabilities, and execution policy. A mode handoff is a state transition—it changes that working context while preserving conversation evidence. The model can request the transition, but the harness performs it. 🔀 Within a turn, the model chooses whether to request tools, hand off, ask for input, or offer a final answer. The harness handles those requests under explicit policies. It can also force a handoff, request intervention, pause, or terminate on its own. Operator commands can pause, stop, or terminate execution. 🔂 These lifecycle outcomes are explicit, typed actions—not just different reasons to break out of a loop. Iteration and tool-call budgets, handoff limits, and stall detection govern when execution must change course. Completion checks distinguish supported success from an unverified claim. Asking for help, exhausting a budget, and finishing the work are distinctly different outcomes. 🧭 Lisa reconstructs workflow progress from its durable conversation log, including plans and completion evidence, so remaining work can be recovered across turns and restarts. Where a mode requires a write, completion checks look for recorded effect receipts rather than accepting the model’s success claim alone. ℹ️ The same loop supports changing roles, multi-step plans, parallel work, recovery, and operator control. The model supplies flexible reasoning; the kernel governs execution and checks completion evidence. So the idea that the turn loop is a simple 3-state FSM is now just a memory; it's gotten a little more complicated over time, @JoshARosen, but à la explicit actions, I'd still call it FSM-based. 😊

This afternoon I gave a junior dev the same advice I have been giving a while: fight it! Especially if you're junior, argue TF out of it. Make that model surrender to *you*, constantly! It needs to be fully answerable to you at all times! The (other) problem with "AGI" is that it puts the tech on a pedestal. It is still just (algorithmic) computers. They are not smarter than you. They're just relentless. Don't surrender, interrogate! 😮‍💨

Lisa Harness v0.5.1 release was last night! 🚀 • Scripted offline e2e testing — zero credentials • lisa-e2e repeat with token/cost roll-ups • Verified-artifact workflow contracts • Incident-driven warning + pack_id fixes • CI smoke gate on every PR github.com/Lisa-Harness/lisa…

Unpopular hot take: Formal verification isn't something one uses *instead of* code review. Its a very stringent technique to ensure that code is mathematically correct and there are non-trivial downsides to using it which will almost always outweigh the benefits. 😒

So, we just do formal verification for app dev now? 😲 I mean, it's a super cool way to make your apps rock-solid, but I got a budget, yo. 💸

I built Lisa Harness on plain-old Python. I used dataclasses and Pydantic, but otherwise it is pure Python from scratch. I didn't even use a DI library for what is ultimately a pluggable framework. It's also 💯 AI-generated. 😟 Here's what I didn't cover in the "introducing Lisa" blog: how I built it. 👇 TBH, at first, I thought it was probably a dead end. I was so pessimistic when I read the terrible Python code the model generated that I almost gave up. But I didn't; I generated and read enough to devise a set of instructions (available in the repo with full revision history) that made the model write unreadable but ultimately correct and testable Python code. 🧑‍🏭 To keep it that way, I used ruff and mypy alongside pytest. The quality gates enforce aggressive function complexity maximums and testing-coverage minimums. There's also vulture for dead code. The model instructions for tests are in a separate file in the repo. 🗃️ I added a direct I/O checker that scans the code for direct reads and writes—which aren't allowed; all file I/O goes through MCP servers like lisa-storage or packs with permission to do file I/O, e.g., coding. 🧪 Then there's "packstest" and "e2etest," which test the discrete logic of packs with mock model responses in the former and agent testing of pack modes and operations with real models (i.e., evals) in the latter. 💡TL;DR: I leaned into testing and verification and deliberately avoided reading the code. The result is for anyone to judge: it's Lisa. 🤖 If you're still reading and interested enough to look, please check the agent testing instructions, *test modules, or the I/O checker especially! 📩 🔗It's Lisa-Harness/lisa on GitHub.

Zero-sum thinking is one of the most counterproductive cognitive biases in business relations. 💀 Almost none of the opportunities the world puts in front of us involve any trade-off, but we humans tend to think they do and, nearly always, invent one that seems real. If you're doing (or not doing) something to "protect your interests," then congratulations, you are almost certainly on the wrong side of this well-known bias.😮‍💨

"You don't need to be well-versed in whatever you're trying to do." 🤨 That's called being a dilettante.😐 Solve problems you understand. Don't waste your time being a dilettante. Internet-famous people get attention for that — you won't. 😒

Oof. I'm not really up on this but I was getting the idea this was becoming an issue in places. From my perspective it seems like a balancing act but if there's no moderating force, it will devolve naturally and, as I type it, I suppose there's some inevitability in that. 😬

In today's episode of "missing the point entirely," its 👇

I'm glad someone says this sh*t out loud. Its rough not having half a Stanford degree or a rich daddy!