Hey, I built a public 7-project tutorial that uses...
# proj-agent
a
Hey, I built a public 7-project tutorial that uses OpenHands for learning harness engineering. I wanted to get your all feedback before I push it wider. Details in the thread.
šŸ™Œ 1
The approach is a hands-on way to learn about how a harness is built. I treat the harness as five levers — model, tools, memory, safety, architecture and have learners change one lever per project against the same baseline prompt, then read the trace. What the tutorial covers: • Model: pre-conversation routing (one concrete LLM handed to Agent). • Tools: lexical (terminal / file_editor) vs MCP search_code — ships a stdlib BM25 server so the experiment is real. • Memory: AGENTS.md + default condenser, plus a trace-diff tool for measuring re-discovery savings. • Safety: policy template + EnsembleSecurityAnalyzer + ConfirmRisky + DockerWorkspace. • Architecture: APIBasedCritic + IterativeRefinementConfig + capstone harness.py. What I'd love your take on: • Any other major levers I should include? • Anything you'd promote from "supporting tool" into a main project — Hooks, Skills, sub-agent delegation, stuck_detection, per-role LLMs (agent vs condenser vs critic)? • Anything obviously missing? https://github.com/rajshah4/learn-openhands-harness
r
Reviewing it now
šŸ™ 1
@average-processor-75195 open to a few isssues/PRs?
a
yea, that would cool, I will take the feedback however you want to provide it. Really appreciate you taking the time to look over it.
r
Your code is excellent. Three lanes I could contribute on: 1) pedagogical doc polish, 2) security module extensions based on the SDK analyzer work I've contributed, and 3) supply chain hardening.
I assume the license will be MIT?
a
Yes. It will MIT and I am sure it's weak on security.
r
No worries, I actually have a project I designed to bootstrap python repos with a decent security posture, CI/CD, pre-commit hooks, project config, etc. Won't be much of a lift and I'm happy to help!
A day or two turn around, sound ok?
I'll open a series of PR's with merge orders attached.
a
Thinking about this, • The pedagogical polish is needed. • The goal is to teach about ways to build the harness, so I think a project that focuses on security would be totally cool. • Is supply chain hardening in terms of my general repo - or is it something to with a harness project?
r
Full supply chain hardening would be over doing it, you're right. I meant just pinning deps to hashes. This is your project, as a former educator I would be grateful for the opportunity to help. It did seem like I was coming in hot, though that wasn't my intention.
I can keep it tight, give it my educators touch, a dash of my security SME, that's it.
a
It's all good. It's a good practice to model in terms of pinning. Where I could use your help is how to tune/customize a harness to be even more secure. There are multiple layers of security and showing people how to work with that would be really useful.
šŸ• 1
r
I can do that. I'll dm you with the details.
ā¤ļø 1
c
Looks awesome! Submitted a license if that's okay with you: https://github.com/rajshah4/learn-openhands-harness/pull/1
ā¤ļø 2