my daily driver lately. got into it for the async harness but it's pretty close on ux too, and bug reports don't disappear into the abyss like codex.
Built a fully featured harness around this! In my testing it's been on par with Codex for most things: subagents, MCP, sandbox, resume, compaction.
Merged the best of Claude Code and Codex - UX, ergonomics, and the bits that actually matter.
Check it out: github.com/viktordanov/uah
Call that advanced Baalistic investigation
in 2026 it genuinely makes me pause, sitting here watching what an 8gb 3060 ti can do today.
i keep thinking about how fast this arrived. six months ago it needed a 24gb card. a year ago, saying a 27b would run on a gaming gpu at all got you laughed out of the room. now it just runs, quietly and most people have no idea.
and it's not slowing down either. whatever this card can do today is the floor, i just needed a second to sit with that.
TFW you thought the model had noticed your Twitter bio of “goblin technologist” and was speaking just to you rather than being generally obsessed with goblins. Feel so betrayed
/cc @goblinodds
lambdamechanic.com/blog/tool…
crates.io/crates/tooltest/0.…
have a go, tell me what you think.
best not to recompile regexes every time you run a tool call, apparently. only a 48x speedup so maybe i'm only half-stupid.
Replying to @aras_p
look, I'm sorry, but the rule is simple:
if you made something 2x faster, you might have done something smart
if you made something 100x faster, you definitely just stopped doing something stupid
the weirdest thing about working with coding agents is that you're constantly asking them to write handover documents so you can flush the context window a dozen times a day. feels like i'm constantly firing them and replacing them with their twin brothers.
quick shoutout to @kerckhove_ts - Sydtest github.com/NorfairKing/sydte… is a really pleasant test framework that encodes all the decisions I would have made anyway: parallelism by default, flake detection, visibility of test times, and even visual test profiling.
the weirdest thing about agentic coding is realising that it's basically programming via opinionated complaining.
i was born for this
Performing the sacred and ancient rite of hunting through the menus of my elderly relatives’ televisions to find and smite the devil that is motion smoothing.
Mark retweeted
Since I got so much feedback on my criticism on MCP I wrote down more thoughts on it. I'm also noticing that there is more an more talk about maybe bypassing MCP recently so this fits right in.
TLDR: agentic loops work best with codegen and MCP is not composable to help here.
considering adding "don't parse XML/HTML with regex" to AGENTS.md
they gotta learn the lore












