Checking my agent-loop articles against the live protocol
Value
2.7
■■■□ of 4
Time
37m
active
Elapsed
1h 34m
wall-clock span
Commits
0
from this session
Cost
$8.92
model spend
Tokens
79k
uncached in, plus out
Tool calls
91
by the agent
Outcomes
9
8 left open
Summary
Written by the summariserRob asked for read-only Sonnet research lanes to check the site's published agent-loop articles against the live protocol and sub-agent extensions, then to propose new articles that would improve the site.
Four read-only research lanes audited the published agent articles, and the assistant spot-checked their main claims. Several articles had wrong counts and routes, one finding about reset rates was reversed, and two live rule pages named a private work context. Rob chose a substitution-list fix for the leak, revisions in two batches, three new articles, and a logged task for a delegation conflict. The assistant fixed and deployed the rules, and seven articles plus a draft note went through a voice pass. Rob then approved publishing everything, so the assistant wrote the three new articles, fixed a failing description-length check, pushed, and confirmed the pages live. The reset-rate article and the delegation conflict remain open.
What came out of it
- Commit [commit] re-synced the published rules from current sources, replaced the retired context7 rule with firecrawl (old URL redirects), and made sync-rules and the just check gate fail on second-context names; deployed and checked on the live pages.
- Protocol reference pages re-synced to current agent-docs (contract 9403d8f, pi harness 75d1801); the build passed with no new substitutions.
- Seven published articles revised in place: the agent loop protocol, the machinery around the loop, lessons from the loop reports, Delete before adding (cap and contract size only), the agent transcript indexer, why not Codex for loops, and one agent context is not enough.
- Three new articles published: setting up Backlog.md for agents, bounding a model request on pi, and just as the only task surface.
- The redirects note was published.
- Commit [commit] was pushed and deployed; the new pages returned 200 and appeared in the feed, sitemap, topic pages and OG card.
- Backlog tasks RKB-0024 to RKB-0027 were closed in local commit [commit], which is not pushed on its own.
- HRN-0325 (which pi agents may delegate) was logged on the agent-estate board and committed locally, not pushed.
- Figure set for the loop-reset refresh was generated into generated-images/loop-reset-refresh/; figure prompts for the new articles are in codex/diagram-new-articles-2026-10.txt.
My agent-loop articles had gone stale. The protocol and the sub-agent extensions had changed since most of the pieces describing them went up, so I asked for a check. Four read-only research passes went through the published articles and reported back. I spot-checked their main claims myself before changing anything.
The check turned up plenty. The pi-side counts were out of date: the loop setup now has nine extensions and twenty-three agent files, where the articles said seven and nineteen. The goal cap had gone from 12,288 bytes to 20,480. Escalation routes had changed, and so had the retry and recovery behaviour. The indexer article was behind as well, with four worker containers instead of three, 21 dashboard tabs instead of 20, and pi compaction events now running.
One finding was the wrong way round. My reset article said implementation lanes finished at about the same rate after the reset. The 6 October assessment shows they went up: lane-worker on the medium setting went from 46% to 65%. The caveat is that "complete" means each lane's own report, and the share the root accepted stayed flat at about 60%. The live article still says the old thing, and I'm waiting for the second check, due 10 October, before I correct it.
Two rule pages also named a work context they shouldn't have. I replaced those names with a substitution list, and the sync and the check now fail if one slips back in.
I revised seven published articles in place and ran them through my voice pass. The drafting agents admitted they hadn't fully read my voice references, so I cut one opinion they had invented. I also wrote three new articles: setting up Backlog.md for agents, bounding a model request on pi, and just as the only task surface. The redirects note went up too. It all went live in one push, and I checked the new pages, the feed and the sitemap. The new articles don't have figures yet.
Still open:
- The reset correction, waiting on the 10 October check.
- A delegation conflict. The machinery article says only one pi agent can delegate, but the escalation agents also carry the sub-agent tool. I've logged it on the agent board, but it isn't pushed, so the article still says the incomplete thing.
- The satori Renovate pull request still fails the OG title-fit test. I still need to hold it with a version rule.
Left open
- RKB-0023 (revising Delete before adding with the first reset assessment) is waiting for the second check due 10 October. The published article still says implementation lanes completed at about the same rate, which the 6 October assessment contradicts.
- The seven-day reopen figures cannot be measured until 11 October.
- RKB-0022 (integrating the loop-reset figure set into the agent series) has its figures generated but not yet placed in the articles.
- HRN-0325 is not pushed, and the machinery article still says only complex-worker may delegate, while loop-pi's escalation agents also carry the subagent tool.
- The three new articles have no figures yet; codex/STYLE-BRIEF.md was missing from this checkout, so the existing loop-reset visual system was reused for the prompts.
- The satori 0.44 Renovate pull request fails the OG title-fit test; the assistant recommended holding it with an allowedVersions rule, and that was not done.
- Local node_modules is out of date (satori 0.34.1 installed against 0.41 locked); CI used the lockfile and passed.
- Deferred ideas: re-measuring the loops, the escalation routes, and Loopwatch, all held until more data exists.