Christian DalSanto

Building Party (YC S24).
Bullish on humans in the loop.

[Christian DalSanto @dalsanto_c]

@slicknet Solid tests + full coverage without crazy overlap will lead to happy & efficient ci/cd

[Christian DalSanto @dalsanto_c]

@michael_chomsky If your harness doesn't carry context to subagents, then just make a skill to have the parent agent spin up a subagent for code review. It should be like 1 line

[Christian DalSanto @dalsanto_c]

One of the best things about building harness-agnostic is when one goes down you just switch over to the next one. No downtime.

Insane that people are all about tokenmaxxing but won't derisk the tools they use. If Codex being down affected you today, take that as a sign.

[Christian DalSanto @dalsanto_c]

@dexhorthy Building is way more than vomiting code. There's an inflection point in which more code actually hurts the product experience. Tokenmaxxing != being productive

[Christian DalSanto @dalsanto_c]

@DavidKPiano Nice! Waiting on a slides/keynote implementation. Claude slides needs a ton of improvement before it's ready for primetime

[Christian DalSanto @dalsanto_c]

@garrytan Surely they must be doing multi tenant. They might be semi-persistent. With proper pod sandboxing that would be most economical, else wouldn't scale at all.

[Christian DalSanto @dalsanto_c]

@bcherny used Opus 5.5 with Lean and got 16 PRs for bugs and race conditions in the Claude Agent SDK. I'm curious how much of the work was deciding what to prove. Could this become a routine pass on a repo, rather than something you need a formal methods expert to set up?

[Christian DalSanto @dalsanto_c]

An agent learning from a bad run sounds useful. I'm less sure when that lesson should become a rule for everyone. A workaround for one repo could be a strange default in another.

[Christian DalSanto @dalsanto_c]

Oh man GPT-6 Sol, Luna, and 5.5 Opus on the same day, right after Grok 4.7 yesterday. 5.5 Opus showing a 5 pt bump over Fable 5.1!

If you're building something, you have to be betting on models getting better with increased velocity or what are you even doing.

[Christian DalSanto @dalsanto_c]

@ChrisGPT What would this do that grok bot or muse couldn't? I wonder if there will be a differentiator

[Christian DalSanto @dalsanto_c]

@garrytan Realtime context is what's missing for all these assistants to feel like magic. Who's working on this?

[Christian DalSanto @dalsanto_c]

@garrytan @capydotai Working in codex or claude code now feels like slow motion. builds should be objective based, not turn based

[Christian DalSanto @dalsanto_c]

@minchoi Delegation is the way to go! How does it determine which Bot Eng to route to?

[Christian DalSanto @dalsanto_c]

I'd rather review an agent session for 5 minutes during the build than discover the wrong assumption in prod a week later.

bring folks in early while it's still cheap

[Christian DalSanto @dalsanto_c]

@somehowali It doesn't have to be. There are tools like buzz, theo, party that make it really easy

[Christian DalSanto @dalsanto_c]

@Gaus450 Are people still using just one? They're all good at different things - use them all

[Christian DalSanto @dalsanto_c]

@hiarun02 Yeah create the plan with fable and code with opencode glm 5.3 flash or deepseek v4.1 flash. review code with a mid model like grok 4.6

[Christian DalSanto @dalsanto_c]

@icanvardar astra was top model the first couple of days - not so great since. Looks like they had too much traffic and had to quantize. Doesn't feel like the same model and we don't use it for code/reasoning tasks at all anymore

[Christian DalSanto @dalsanto_c]

@EnvolDev @prasenx You can get a better handle on the quality of your code without looking at it. Try asking the agent more questions before/during/after the build.
Also yeah, tests are super important (unit, component, playwright, etc).