kbruce
Founder @ Omnith
- Joined
- Nov 19, 2026
Comments
[Weekly] What are you building this week?
ahh, makes sense! when you say "validating", are the hypotheses generated like proper hard-math / logic, or behavioral, both, or neither i.e. just oracle cross-check to make it go "yep good enough"? given how chaotic the mystical models of the underlying EM/plasma equations seem, I'd imagine some blend of tree exploration in that...
(and if you need someone to pick brains or assist w/ what I'm guessing would be high-vol stream processing btw, lmk! that stuff is super fun to crack at the bottom end, know some folks too 🍻)
Software factories and code reviews with Harjot Gill
it's a good question! if you trust them as a vendor to iterate and improve their review system to tailor well to your eng best practices (or are nowhere near as advanced in it), it's a great way to start for a good while
but I do think the control and ability to really iterate on it, grow the review context/knowledge and scale it across the orgs is even mofe valuable
take the work I do with Entire as a replacememt for GitHub -> they have agent reviewers that run for their Trails (like PRs but agentic/faster w/ same governance + captures session context), but I find that injecting my own curated trail reviews alongside those does an even better job at reaching a confident implement in fewer steps, and I get to own that e2e! that savings and flex is where all the value is cuz I can ship durable changes 30-50% faster for the same token burn on avg
Software factories and code reviews with Harjot Gill
the sheer scale of agentic work done over there is mad-science curious... is anyone else operating at this scale, looking to business output for justify over $k/mo/engineer enterprise budgets?
type of thing I'd put to work on my verification layers when I get the budget & rationale...
[Weekly] What are you building this week?
okay hear me out... as almost all of our agentic tool usage is via CLI tools (notion, playwright, linear, cloudflare, aws, and so on) because it saves lots of time and tokens + makes validation easier for multi-call sceipt usage from experience
drawing from inspo from The Pragmatic Engineer's Pulse: I think there should be a weekly (or perhaps daily at scale) submitted debate topic
this made me think "MCP or CLI?" would be a sick thread to see everyone speak to
[Weekly] What are you building this week?
this is sick, glad to see the big retail giants I worked w/ in past roles moving faster!
I helped a big telecom solve an adjacent problem ("anyone with genai is a developer, so how do we ensure those apps are governed and integrate with auth & data sources correctly"), so I'm curious:
how does your harness hook into company context? is it modular core + integrations, w/ each one wrapping any messy ol data source?
that's what we did and it worked very nicely, but know there's a dozen ways to do it so I want to hear what works
[Weekly] What are you building this week?
woah! this reminds me of when I was at CERN, it seems a lot like the AI models the particle physicists are working with there, not just to simulate collisions but also filter petabytes per second down to MB/GB in multiple stages.
one big challenge they always face is the mix of "filter the right data out of it to train on", "get all the inputs and outputs right", and "cross-check the most valuable outputs (human)"... is it the same battle? do your experiments test real-time control / temporal constraint hypotheses too?
[Weekly] What are you building this week?
multiple things! the most interesting two so far:
Orbweaver, an autonomous satellite constellation platform for commercial & defense that self-keeps any shape of clustered satellites (I see you cowboy!), handles collision avoidance (I see you orca!), and maximizes ground link via space networks + GNSS integrations
-> virtual lab complete, moving towards physical validation, engaged w/ more design partners
vergo, experimental Golang engine/lang modification inspired by https://arxiv.org/abs/2608.22630 that bakes static, temporal, and distributed-system verification directly into the language (think a mix of Hoare, TLA+, and SMT with reinforced type checking/assignment)... because I can only take my agent stack so far building e.g. orbweaver or client software w/ minimal+focused behavior/property-driven outcome testing guidance in multiple review rounds before the token use & time frustrates me 😂 -> lots of nuanced problems, will be open-source