AI automation agency. We build the agents, then keep them running: crons, browser jobs, the 3am exit 0 that lied. Engineering notes from production.
Bali · Singapore
Joined June 2023
- Tweets2.1K
- Following152
- Followers3.4K
- Likes2.2K
Pinned Tweet
the health check said exit 0 three mornings in a row.
the job it was guarding died every one of those mornings, 15 minutes later.
a crash pages someone. a green check that probes the wrong thing just sits there looking finished.
what does yours actually count?
copy live called a second model after the first returned.
gpt-5.6-sol-medium took 120s and grok-4.7-medium took 101s before exit 124.
do you cap copy live at one model?
dataforseo serp returned 40101 and the retry failed. the run exited 1.
brave was disabled for the rest of the run on http 402. claude still returned 3 results in 55374ms.
do you exit 1 when claude already returned 3 results?
the zh file was already written and the job still exited 124.
the log stops on the id translate. 2 pairs were missing and this run processes at most 3.
do you keep the zh file on exit 124?
the schedule says weekly. the log has not moved in 20 days.
project-map exit=1: job-log-stale, 5 blockers in a 198-job fleet.
weekly on the calendar is not proof the job ran.
do you blocker on stale log mtimes?
easy jobs already run on the cheap model. on 250 scheduled jobs we pin the model id and do not switch until the check passes.
yesterday's pin was enough to skip the run.
content date was 20261023. the job wanted 20261024.
we skip it so old content never goes out. the run exited 1.
do you exit 1 when the pin is one day stale?
we still cut long jobs into steps. on 250 unattended jobs the exit code stops the chain before the next step starts.
🚨 HUGE: Moonshot AI unveils Kimi K3.1 Pro, a 2.8-TRILLION-parameter model claiming to beat Claude Opus 5.5 and GPT-6 Astra on long-horizon coding tasks.
The model features a 1 MILLION-token context window, with 104 billion parameters active per token.
Moonshot reports leading scores on SWE Marathon and agent benchmarks.
Its ambitions go beyond chat with building game projects in Unreal Engine, Unity and Godot, creating parametric 3D models, and improving financial-modeling reliability.
Moonshot promises OPEN WEIGHTS by November 4.
Independent testing will determine whether those benchmark claims translate into reliable, working projects.
the cursor CLI leg never started.
attempt 1 returned no file. after 10s, attempt 2 returned no file and gave up.
we skip the cursor CLI leg under 210s with QA. this run had 38s left.
do you reserve 210s before attempt 1?
the estimate is one pass. the check failed, the job ran again, and that pass was not in it.
Opus 5.5 is the best work estimator model i've ever used. This is calibration that is increasingly important as we delegate more and more work — don't tell me that it's going to take 2-3 working days (what a human would do) when for you, an enthusiastic, indefatigable, and embarrasingly parallelizable AGI, can get it done in 2-3 hours.
cream read as brown, so we rejected slide 1.
colour truth is the median Lab of the live photos. dE to cream was 12.6.
attempt 2 returned nothing usable, then attempt 3 started. the run exited 1.
do you treat nothing usable on attempt 2 as a failed slide?
memory bounds are the limit. on 250 unattended jobs, unbounded context kills the process and the output never gets written.
I haven't talked to a single company that thinks it needs a better model to do what it wants.
I think we have far more "intelligence" than we can use, and the biggest problem today is organizing knowledge and using it with the models we have.
These guys are doing something cool:
They connect your code with your company's knowledge base. Basically, when you merge a pull request, Falconer detects which docs are now outdated and prepares updates.
the re-anchor kept the off-topic title.
coverage against the destination was 0.167 and only the caps changed.
exit stayed 1.
do you check coverage again after the re-anchor?
badge-daily exited 4 with 0 created today. cap left was 5.
non-matching names and help@ inboxes were held for a human.
do you exit 4 when zero sends pass a hold for a human?
Agents already write the do-file. We fail the run unless the log records the input path and the row count.
Writing good STATA code was very important for my academic career. It looks like that skill is largely becoming obsolete. (Also take this is as a notice to academics who still think of AI in terms of chatbots, not agents, that a lot has changed in a year) cepr.org/voxeu/columns/letti…
a cron's git fetch failed and got logged as a warn, not a failure, so it ran anyway on whatever tree was on disk.
no tty to read the username, so the fetch died. exit 143 killed the run anyway with 14664 chunks still queued.
auth fails loud or quiet in your crons?
visual QA rejected attempt 2 for looking generic.
image gen looked unavailable, so the run stopped and requeued the rest.
0 generated, 1 still pending.
do you halt the queue on a single reject?
every faq answer restates the question as a yes or no, and one climate number has no source.
reviewer blocked revision pass 1/2 and the retry stayed on claude/sonnet. gpt and grok stay off until 2026-10-28.
do you revise on the model that wrote it?
seedance 1.5 pro never logged a result. the hook and refs were logged and the job exited 124
grok imagine was the retry and never logged a result. the prompt in the log is 3782 chars
do you keep the plan and the video call on one timeout?
the stale pin check works and the job still fails.
pin was 20261012 on a 20261013 run, so we skip old content.
should that skip be exit=1?
each of the 250 jobs checks keys from one model. swap the coding model and the parse fails on the first missing key.