kyle (@kyleryan) header
kyle (@kyleryan) profile picture

SI agent manager. trying to keep up with the singularity

Joined March 2021
Photo by kyle (@kyleryan)Photo by kyle (@kyleryan)Photo by kyle (@kyleryan)Photo by kyle (@kyleryan)Photo by kyle (@kyleryan)Photo by kyle (@kyleryan)Photo by kyle (@kyleryan)Photo by kyle (@kyleryan)Photo by kyle (@kyleryan)Photo by kyle (@kyleryan)
every AI chat ends the same way. you pasting the error back in and typing still broken
13
if you're trying iphone duo in claude code, grep for hardcoded screen sizes first. the demo app drew on a fixed phone-size canvas and sat in a tiny box mid-screen then make claude test closed, half open and open, and fold mid-task the fold is the new "works on my machine"
2
2
200
everything i work on is so trivial that Fable 5.2 and Opus 5.5 high feel exactly the same to me
108
still trusting one good agent run? that proves it can, not that it will vercel's evals guide: - pass@k: any of k runs passes. fine when retries are cheap (coding + tests) - pass^k: all k pass. use when a miss costs the user (orders, payments) 5/5 just means no failures seen yet
1
3
126
a long system prompt is tech debt the model pays interest on every single turn
2
2
87
codex capacity errors fell from 6.82% of daily users to 0.085%, and every fix marked on this chart is retry timing if your agent loop hits overloads: - read retry-after - back off with jitter - cap retries, then stop otherwise your while loop is the capacity error
2
98
GitHub says some teams are having trouble merging code this morning if your merge times out, refresh to see whether it went through before retrying. monday has enough duplicate work
1
1
226
at 7:35 a.m. ET, GitHub said about 2–3% of merges were timing out, up from about 0.2% normally. it’s still investigating githubstatus.com/incidents/2…
20
before trying to catch Ship 42, SpaceX checked its heat shield with cameras on satellites the ship had just deployed
Mechazilla catches Starship! Congratulations to the entire SpaceX team on this incredible achievement!
104
they are about to catch a skyscraper that just orbited around Earth with toothpicks
89
the FAA ground stop is a good day to see what your flight app means by “on time”
1
217
parallel agents don't stall on code. they stall waiting on you claude code's new projects demo has a "waiting on you" lane. the rate limiter thread stops to ask redis or in-memory. make the boring calls first: - where state lives - which deps are ok - what done means
1
2
128
Argon, Carbon, Hydrogen what is any of this supposed to mean? can we please just go back to small, medium and large is Argon to Terra what Luna is to Hydrogen? or is Hydrogen the Terra one? is Argon the Luna model? what am i even doing here
117
most llm-judge evals are a vibe check with a spreadsheet anthropic's eval guide, ported to agent evals: - blind hand-label a sample, check the judge agrees - one bad turn fails the whole chat - pick the metric before you look - single-turn is saturated (94-98% on recent claudes)
7
3
207
switched models? your claude.md is still prompting the old one run /doctor prompt-audit: - reads claude.md, agents.md, skills - flags old-model habits, dead paths, contradictions - proposes diffs, edits nothing till you say so it skips settings, so hooks are still your problem
1
1
90
every data team has that one query nobody touches. 600 lines. no comments. the person who wrote it left in 2021
91
people who started agents on opus 5.5 skipped the free distrust training the bad models gave the rest of us steal it anyway: - ask what it skipped, not what it did - make it run the tests before it says done - commit before every big ask the magic is real. so is the rollback
53
the Outlook outage is a good time to ask what your agent does when it can’t read the inbox “no urgent email” and “couldn’t check” need different states. otherwise the agent has discovered a very efficient way to achieve inbox zero
1
172
set your agent's stop list before you set its personality chatgpt dots now follow custom rules, including which actions need your go-ahead. write those first: - anything that sends - anything that spends - anything that cancels "always talk in lowercase" can wait
2
2
111
most hallucinated api calls are your stale readme being followed perfectly. fix the docs before you blame the model
1
93