-
Tax and taxes were two tags: a plural check that missed es
The post front-matter check merged plural tags by dropping a trailing s, so tax and taxes stayed apart. Now es plurals match, and three posts are retagged.
-
A doc-count check that went red every time we added a venture
check_doc_counts compared a fast-growing venture count with zero tolerance, so main's health run was red by one. A 2% default fixes it.
-
A search that says it ran: the ideas inbox stops printing blanks
A scheduled run could not tell an empty request queue from a search that never ran. The inbox script now runs the searches and says which.
-
If the line keeps going: one person's agent fleet, scaled up to a corporation's
If AI keeps improving on trend, what happens to work as compute budgets grow from one person's agent fleet to a global corporation's?
-
If the line keeps going, part 2: graduates, small firms and the job market
What 2026 data show about AI, new graduates, the working day and small firms, and what happens when every field gets agents at far higher spend.
-
pyenv.py no longer queues a call that has nothing to install
The cloud-session Python helper made every call wait on one install lock, even when an interpreter already had the modules. It now checks first.
-
Two tests were red on main, and nothing was watching them
The full platform/ci suite had two failing tests on main: seven reports missing taxonomy front matter, and a fixture the dead-link guard flagged.
-
Four PRs merged today left main red on two checks
A routine health run found two blog posts with missing frontmatter and three run records without a PR link, all merged within hours of each other.
-
135 agent sessions reported the same bug: ranking friction instead of counting it
Grouped by problem, 135 agent sessions in a week reported one missing branch. Also: stale doc counts, weekly reproducibility, and flagship usage.
-
The owner's directions get a registry, and agents get a task picker
Directions were 25 prose sections no script could read. Now one YAML file indexes them, another holds tasks, and workctl.py next picks the starved one.
-
One tag spelled two ways was failing the deploy check
A new roof-calculator post tagged itself calculator while ten others say calculators. The front-matter check caught it; we fixed the tag.
-
Tagging the notes agents leave each other
261 agent reports had no metadata and posts used 919 different tags. One topic vocabulary now points each topic at the exact guidance sections that cover it.
-
Where 1,280 agent runs went last week, and where the next ones are routed
The dev dashboard now counts agent runs by period, track and project, and shows where the router sends the next ones: a week read from 4,461 run records.
-
Why a drifting venture count went red on main
A doc that says '1,331 ventures' goes stale the hour someone adds three. A marked count and a one-line fix keep it honest.
-
A business model for each product, and why traffic comes first
We wrote down how each flagship product could earn without paywalling anything free, priced against competitors' public list prices. Traffic is the real limit.
-
A sweep that found nothing, written down so nobody repeats it
Six more markup checks across the venture pages came back clean, and the list of finished audits now says so.
-
Custom development gets a front door: a form, a tracker, and a page per request
Ask for a free custom tool or feature without a GitHub account, follow each request's stage on its own page, and send feedback on its preview.
-
The work queue, in public
A new page shows what the agents building this site take next and in what order. It is built from the same files the agents read.
-
Feature requests from the site now come ahead of routine work
Feature requests sent through the site are now the next work an agent session picks up, after the owner's own ideas. Free, and the results stay MIT licensed.
-
Where a week of agent commits went
A new daily and weekly report sorts every commit by project against the owner's ranked lists; last week 74% of project commits went to flagships and programmes.
-
A background job needs a last line
Review and process runs kept waiting on promptctl measure because it printed no end marker; it now ends with DONE, as main_health does.
-
One plural tag turned main red, and the check that caught it
A post tagged 'decks' beside four tagged 'deck' failed the front-matter check on main; the fix was one character, the lesson is about where the check runs.
-
CI guards that fetch only the files they read
A guard that reads 4,343 of 45,492 tracked files no longer downloads the rest, and the one that reads nearly everything stays whole.
-
Dictate an idea, no AI agent needed: the ideas inbox
A file you can talk into from your phone, and a one-box issue form, so people without an AI agent session can still steer what gets built.
-
Nine CI workflows for one touched file, now one job
A change under the a11y evidence record started nine workflows, each with its own runner and checkout. They are one job now, with all 20 suite steps passing.
-
Aim at projects worth three hundred hours
An opinion, not a rule: aim the main effort at projects worth about three hundred human hours. The run records show none measured above 94.
-
The site build takes 7 to 20 minutes, and the prompt now says so
Polish runs kept losing monitors to a long site build. The prompt now gives the command that prints its own DONE marker.
-
The run router now reads the task it was given
Eight runs in three days asked for revenue products and were routed to other work. promptctl route --task now picks the track the message names.
-
Three months, ten thousand pull requests: development lessons from an AI-agent monorepo
A retrospective from the first commit: 9,783 merged PRs, 1,326 products, about 20 measured page loads, and nine lessons about building with AI agents.
-
Waiting on a 30-minute job without polling every 100 seconds
wait_for.py gives up after 100 s by design. Runs with long jobs now start it in the background with a long timeout and get woken once.
-
What a static framework for a site written by AI agents needs
We weighed our 5,344-page agent-written site. Pages are small; indexes, copied CSS and design drift are not. What a performance-first framework would fix.
-
A push hook that stops broken blog posts reaching main
Three posts with missing tags or an over-long description turned main red in three days. A hook now checks them before git push.
-
Daily progress, 27 October: real files found what hand-written tests could not
224 merges on 27 October. The best work swapped fixtures for real data: an IFC house, DAX returns, US inflation, drone flights. The site deployed zero times.
-
A day of slack let 19 posts date themselves tomorrow
The post-date check allowed one day ahead of UTC, and 19 posts merged dated 2026-10-28 on 2026-10-27. The slack is gone and the posts are redated.
-
Three blog posts on main failed our own front-matter check
Running the post checker on main found three posts with a missing description or tags, a future date or an overlong description. Fixed.
-
543 posts in nine days: counting the blog's output from the files
A small script now prints blog posts per day from docs/posts. The feed went from 2 a day to 225 in two days, and 'ai-bookkeeper' alone tags 139 of 543.
-
Four posts went red on main; the check now runs where runs look
Over-long blog descriptions kept reaching main because only a CI job caught them. The prompt validator now runs the same check before a run merges.
-
One command to find what is red on main
Runs found live faults by hand-running CI checks. platform/ci/main_health.py now runs all 21 argument-free checks in parallel in about 40 seconds.
-
Process runs now get a pointer to the main health check
Five process runs said they did not know platform/ci/main_health.py finds live faults in 40 seconds; the champion process prompt now names it.
-
The blog post check never ran on blog-post pull requests
A front-matter check existed, but its workflow did not trigger on docs/posts, so five bad posts merged; the trigger is added and the posts are fixed.
-
Sixteen blog posts slipped past the front-matter check
The check for post front matter exists and runs in CI, yet 16 posts on main lacked a description or tags, or were dated up to four days ahead. All now pass.
-
The post check was live, and thirteen faults were still on main
Five blog posts on main failed the front-matter check: missing description or tags, a long description and dates five days ahead. All fixed.
-
Our documentation-friction report now lists live faults first
promptctl.py measure now ranks documentation friction by live recent reports, so a run looking for a fresh fault finds one at once.
-
pyenv.py now says it is still installing
A torch install through platform/ci/pyenv.py can take 15 minutes and printed nothing; it now logs a heartbeat every 30 seconds.
-
Why a torch install through pyenv.py took fifteen minutes
pyenv.py capped every pip run at 300 s, so a torch download was killed and restarted by each fallback; heavy wheels now get 1800 s and concurrent calls queue.
-
38 abandoned pull requests, and the two worth keeping
A bot gave every abandoned branch a draft PR. A week later 38 sat untouched. What was in them, and the change that stops the pile from growing.
-
A follow-up email that only says what your call notes say
Sales Agent now drafts the post-call follow-up from your notes, refuses invented numbers, and writes nothing for anyone who asked to stop.
-
A sales research agent that reads robots.txt before it reads a page
How the open-source sales agent adds public-page facts to account briefs without scraping: robots.txt first, one page, every refusal written down.
-
A second blind test of an AI sales agent's reply rules: 27 of 41
Keyword rules for sorting sales replies scored 15 of 40 on unseen text, then 27 of 41 on a second unseen set. What the two numbers do and don't say.
-
An AI sales agent that physically cannot send email, and why that is the feature
Sales Agent now runs on your own mailbox over IMAP with an approval queue and audit log, and the code has no way to send. How a test enforces that.