Skip to content

Loop Engineering News Sources

The fetch-loop-news skill reads this file on every run. Add, remove, or edit rows here to change what gets tracked — no changes to the skill needed.


Relevance keywords

Scored in four tiers. A post matching Tier 1 or 2 is always included in the digest. A post matching Tier 3 or 4 only is included if the subagent judges it substantively relevant to loop engineering practice (not just a passing mention).

Tier 1 — Boris Cherny's exact language

Any post using these terms is directly on-topic.

  • "write loops" / "writing loops"
  • "my job is to write loops"
  • routines (in a Claude Code / agentic context)
  • "loops that prompt"
  • "the thing that writes the code"

Tier 2 — Named discipline (Osmani / Steinberger framing, built on Cherny)

Widely adopted community terminology.

  • "loop engineering"
  • "replace yourself as the person who prompts"
  • "designing loops"
  • "agent loop"
  • "agent harness" / "harness engineering"
  • "factory model" (AI software factory)

Tier 3 — Named concepts within the space

Specific patterns and failure modes; include if the post discusses them in an agentic context.

  • worktree (isolated git checkout per agent)
  • subagent / sub-agent
  • "maker checker" / "writer reviewer"
  • "intent debt"
  • "comprehension debt"
  • "cognitive surrender"
  • "orchestration tax"
  • "adversarial code review"
  • "verification loop"
  • "stopping condition" (+ agent / loop context)
  • headless (+ Claude / agent)
  • MCP (+ agent / connector)

Tier 4 — Tool and feature names

Surface on-topic content; include only if the post discusses agentic / loop patterns, not just general usage.

  • Claude Code
  • /goal (Claude Code command)
  • "permission mode auto"
  • Codex (OpenAI, agentic context)
  • OpenClaw
  • agentic / multi-agent
  • "tool use" (+ agent context)

Sources

A source row that returns nothing looks identical to a quiet week. The Batch row below was configured as rss against a feed that had 404'd. For two months it produced zero findings and nothing flagged it — because "no new posts" and "the feed is gone" are the same empty result. It cost the KB an entire four-part Andrew Ng series that turned out to be the spine of Part II.

A source yielding nothing for more than ~3 consecutive runs must be re-fetched by hand before its silence is believed. Rows carrying a 404 note are in that state now: treat them as unswept, not as quiet.

Revalidated 2026-09-06 00:26 UTC. Every one of the 47 checkable URLs in this table was fetched. 45 returned 200. One was dead — AI Breakfast, fixed below by the same rss → html switch The Batch needed, because no discoverable feed exists. One stored an http:// URL that 301-redirects; it now stores the https:// form. The remaining 7 rows carry handles or search queries rather than URLs and are not URL-checkable. No other row is broken.

Actor Type Handle / URL Notes
Anthropic html https://claude.com/resources/articles Primary source for Claude Code updates. No RSS feed exists (confirmed Jul 2026 — /rss.xml, /news, /blog/rss.xml all 404, no <link rel=alternate> in page head); official blog now lives at claude.com/resources/articles; claude.com/blog 308-redirects there (checked 2026-10-19). The index renders only 11 of ~285 posts ('View more' is client-side), so also read the category pages (/resources/product-announcements, /perspectives, /best-practices, /articles?page=2): the HTML embeds Sanity JSON ("_type":"blogPost": date, slug, externalUrl). Index dates can disagree with the destination page for externally hosted posts (Oct 16 vs Oct 15; Oct 7 vs Sep 28; Oct 19 vs Oct 13) — use the destination page date. /blog/rss.xml still 404 (re-checked 2026-10-19)
Anthropic html https://www.anthropic.com/news Distinct from the claude.com/resources/articles (formerly claude.com/blog) row above — a different domain (anthropic.com vs claude.com), not tracked until now. Added 2026-09-28 after a first-party Anthropic post, the Sep 23, 2026 novel-enzyme-discovery announcement, reached the digest only secondhand: it is not on claude.com/blog (confirmed — no mention on that page), so this dedicated newsroom URL went unswept. www.anthropic.com/news itself returns 200 (checked 2026-09-28); the /news 404 noted in the row above was checked against the claude.com domain, not this one 2026-10-19: no RSS/Atom (/feed, /rss.xml, /news/rss.xml all 404); curl with a browser User-Agent returns 200 and the page embeds JSON (publishedOn, slug, title, summary) for ~100 posts — parse that and pair slug/title per object (a naive regex mispairs them). Low-yield for loop engineering: 10 posts in 2026-09-28..10-19, none Tier 1/2 (policy, partnerships, safety)
Anthropic (developer blog) html https://claude.dev/blog/ Anthropic developer blog; carried 6 of the 2026-10-19 run's findings (Osmani, Lance Martin, Thariq Shihipar: code review, automations, cloud sessions, evals). Server-rendered index with dates and authors; LOAD MORE is client-side, so only the ~10 latest posts are visible. Index dates can disagree with the article page for externally hosted posts (code-review Oct 16 vs Oct 15; eval/hillclimb Oct 7 vs Sep 28) — use the destination page date (added 2026-10-19)
Anthropic Engineering html https://www.anthropic.com/engineering Anthropic Engineering blog. Added 2026-10-19: repeatedly cited but not tracked as its own row — grep -c 'anthropic.com/engineering' docs/*.md finds links in 12 docs files (19 matching lines), so new posts only surfaced secondhand. Trigger: the Scion/source-control post (2026-10-13; the claude.com index showed it as Oct 19, so use the destination page date)
Boris Cherny x @bcherny Creator of Claude Code; coined "write loops" Logged-out X (checked 2026-10-19): search sits behind a login wall; the profile shows ~8 posts (newest 2026-10-16, oldest 10-07) then a 'Continue to X' gate, with no time[datetime] (dates are text like '16 Oct') — confirm post dates via api.fxtwitter.com//status/
Andrej Karpathy x @karpathy Influential ML researcher
Andrew Ng x @AndrewYNg Agentic AI education
The Batch — Ng's letters (DeepLearning.AI) html https://www.deeplearning.ai/the-batch/tag/letters High value; was silently unswept for two months. Andrew Ng's weekly letter. Read the OPENING SECTION of every letter regardless of its title — the coding-agent material is usually at the start, so a letter titled about export controls or course design still carries concrete agentic-coding guidance. Filtering by title loses most of the signal. Switched rss → html on 2026-09-04: the official feed (/the-batch/feed/) has been 404 since at least 2026-07-08 and no <link rel=alternate> exists, so the RSS row silently returned nothing for two months while looking configured. The tag page scrapes fine and paginates. A 2026-09-04 sweep of 57 letters found 37 with on-topic content, including the four-part AI Engineering Skills Map whose "Using Coding Agents" letter is the spine of KB Part II
OpenAI rss https://openai.com/news/rss.xml Feed discovered Jul 2026 (was html-scraped; 250+ articles returned, root-level path works despite prior 403 note) 2026-10-19: feed healthy (200, 1299 items) but carries only title/description/link/guid/pubDate, and every openai.com/index/ page returns 403 to curl and WebFetch (no Wayback snapshot) — findings rest on feed title/description only
Addy Osmani rss https://addyo.substack.com/feed Corrected Jul 2026 — his active blog is the "Elevate" Substack (addyo.substack.com), which carries the origin "loop engineering" post (Jun 2026); the previously-tracked addyosmani.com feed was a different, largely-dormant site
Simon Willison rss https://simonwillison.net/atom/entries/ LLM tooling practitioner 2026-10-19: this feed carries only long-form posts; link-blog items, quotes, notes and releases are only in /atom/everything/ (30 entries, 2026-10-07..10-18) — consider switching, expect higher volume. Neither feed reaches back more than ~12 days, so for longer windows read the /YYYY/Mon/DD/ archive pages
Swyx x @swyx AI engineering community. Low yield Jul 2026, confirmed again with a broadened query — his own posts are consistently off-topic reposts; findings so far have come only from reposts/threads, not his own writing. Keep tracking (profile scan is cheap) but do not expect direct-post yield
swyx.io rss https://www.swyx.io/rss.xml AI engineering long-form posts; feed URL discovered Jul 2026 (was html-scraped). Feed is alive (200 OK) but most recent post is dated 2026-07-05 — no new content in ~2 months as of 2026-09-06; low priority until it resumes posting Checked 2026-10-19: HTTP 200, newest item 2026-07-05, nothing since 2026-09-28 — dormant ~3.5 months; keep low priority
The New Stack rss https://thenewstack.io/feed/ Active loop engineering coverage. WebFetch on thenewstack.io returns the newsletter-signup wrapper, not the article body — use curl + tag-stripping for full-text reads of this source (confirmed Sep 2026) 2026-10-19: the feed returns only the newest 25 items (~4 days), so a ~3-week window needs ?paged=2..N (paged=2..15 gave 26 items each, back to 2026-08-14). Items carry full text in content:encoded, so curl on the feed suffices. Repeats sponsored/vendor-authored pieces (Signadot, CoreWeave, Aviator) — flag them as vendor-authored
Sabrina Ramonov rss https://www.sabrina.dev/feed Loop engineering + /goal + Routines (Jun 2026); feed URL discovered Jul 2026 (was html-scraped)
Sabrina Ramonov x @Sabrina_Ramonov Active X presence posting loop engineering definitions and techniques (Jun 2026)
Cobus Greyling rss https://cobusgreyling.substack.com/feed Loop Engineering Substack + GitHub repo
Peter Steinberger x @steipete Creator of OpenClaw; "designing loops" framing
Lenny's Newsletter (Claire Vo) rss https://www.lennysnewsletter.com/feed Four-type loop taxonomy article; high-engagement AI practitioner audience (Jun 2026); Substack RSS Checked 2026-10-19: OK (20 items back to 2026-09-26); bodies are mostly show notes or free previews, and paid episodes expose only the outline, so claims from them are bullet-level
Cobus Greyling loop-engineering github https://github.com/cobusgreyling/loop-engineering Reference implementation and patterns repo; watch for new docs, examples, and releases
The Rundown AI rss https://rss.beehiiv.com/feeds/2R3C6Bt5wj.xml Daily AI newsletter (Beehiiv); scan for loop engineering / agentic workflow coverage. Feed URL rediscovered Jul 2026 (original /feed 404'd)
TLDR AI rss https://tldr.tech/api/rss/ai Daily AI digest RSS; scan for Claude Code, agent loop, agentic workflow coverage. Feed URL rediscovered Jul 2026 (/ai/rss 404'd; tldr.tech/rss is the general-tech feed, not AI-specific) 2026-10-19: items carry ONLY title/link/pubDate (emoji teaser titles), so keyword scoring on the feed misses everything — fetch each issue page (https://tldr.tech/ai/YYYY-MM-DD, curl -sL) in the window and ignore the '(Sponsor)' blocks
Ben's Bites rss https://bensbites.com/feed Daily AI news digest RSS; scan for loop engineering / agent loop coverage. Feed URL rediscovered Jul 2026 (moved off beehiiv to its own domain)
AI Breakfast html https://aibreakfast.beehiiv.com/ Daily AI newsletter; scan for agentic workflow and Claude Code coverage. Switched rss → html on 2026-09-06 after the same failure as The Batch: /feed has 404'd since at least 2026-07-08, /feed.xml and /rss also 404, rss.beehiiv.com/feeds/aibreakfast.xml 404s, and the homepage carries no <link rel=alternate> and no beehiiv feed id — so no discoverable feed exists. The index page resolves 200 and lists ~9 posts as /p/<slug> links; scrape it directly. The row's previous note said "try /rss next run" and sat unactioned for two months 2026-10-19: the index showed 9 posts (09-30..10-19) but /archive (200) lists 12, adding 09-23, 09-26, 09-28 — use /archive for ~3-week windows. Post pages carry datePublished in page JSON plus full text, so curl works; issues run ~Mon/Wed/Fri; high-yield for Claude Code release roundups
X search — loop engineering x-search https://x.com/search?q=loop%20engineering&src=typed_query&f=live Live keyword search; dynamically loaded — scroll ≥3 times to surface 20+ posts. from:<handle> queries drop their author scope when the OR-clause isn't parenthesized, returning unrelated global results — wrap it: from:<handle> ("term1" OR "term2" ...) (confirmed independently by multiple sweep agents, Sep 2026) 2026-10-19: the sweep's Chrome profile was not logged in to X — login wall on /search, so Pass 1 and Pass 2 are unreachable until it is
LinkedIn search — loop engineering linkedin https://www.linkedin.com/search/results/content/?keywords=%22loop%20engineering%22%20OR%20%22harness%20engineering%22 Professional community posts; dynamically loaded — scroll ≥3 times to surface 20+ posts. Keywords broadened 2026-09-06 from loop engineering alone: adding harness engineering is what surfaced 2 of the 3 substantive posts in that day's baseline 2026-10-19: the default 'Latest' sort (sortBy=date_posted) returns only same-day noise; adding datePosted="past-month" with sortBy="relevance" reached 68 posts in ~36 scrolls (not date-ordered; ~10% recruiter/unrelated hits). Post permalinks are mostly absent from the DOM (only /pulse/ links), so non-article findings can cite only an author profile; Pulse pages fetch with curl and a browser User-Agent
Harness Books (AgentWay) html https://harness-books.agentway.dev Essay collection on harness design theory — unstable components, ledger closure, input governance, reactive compact Checked 2026-10-19: static, undated, no RSS (/feed, /rss.xml, /atom.xml, /feed.xml, /search all 404); last modified 2026-04-19 and wquguru/harness-books last pushed 2026-04-19 — nothing new; low-yield, consider monitoring that repo's commits instead
MindStudio Blog rss https://www.mindstudio.ai/rss.xml Published 3+ loop engineering articles covering loop design, agent harness architecture, and verification patterns (Jun 2026); feed URL discovered Jul 2026 (root-level path — /blog/rss.xml 404s)
GitHub search — loop engineering claude github-search https://api.github.com/search/repositories?q=%22loop+engineering%22+claude+stars:%3E3&sort=updated&per_page=100 Search GitHub repos combining "loop engineering" + claude; returns JSON — no auth needed (10 req/hr limit); find new implementations. 179 total results, heavily diluted by same-day 0-star template/farm repos — add a stars floor (e.g. stars:>3) to cut noise (Sep 2026) Stars floor adopted 2026-10-19 (stars:>3, per_page=100): 38 results vs 179+ unfiltered, no farm noise. Do not drop the floor — a created:>2026-09-28 query without it returned 36 repos, 28 of them 0-star loop.js forks (a fork farm). sort=updated is a weak recency signal (updated_at moves on star/watch/fork events): check pushed_at, not updated_at. Complementary query to try: q=%22loop+engineering%22+pushed:%3E2026-09-28+stars:%3E3&sort=updated. Keyword saturated by skill/template repos: expect few finds
GitHub search — claude code harness github-search https://api.github.com/search/repositories?q=%22claude+code%22+harness+loop+stars:%3E5&sort=updated&per_page=50 Search GitHub repos combining "claude code" + harness; returns JSON — surface new harness design repos. 4,529 total results as of 2026-09-21, still heavily diluted — narrow to q="claude code" harness loop or q="claude code" harness orchestration (was orchestrat; that stem returns 0 results — corrected 2026-10-19), and/or add a stars:>5 floor Narrowed 2026-10-19, works: "claude code" harness stars:>5 -> 597 results; + orchestration -> 47; + loop and pushed:>2026-09-28 -> 14. Pitfalls: sort=updated is dominated by same-day pushes, so the pushed:>YYYY-MM-DD filter is more useful than the sort; the unauthenticated limit (60/h) was hit mid-run, so send a token or keep to 1-2 search calls; a DeepSeek Harness ('dsh') plugin cluster dilutes the broad query (exclude with NOT deepseek). Yield ~5 new Tier 1/2 repos: keep the row
cobusgreyling/goal-engineering github https://github.com/cobusgreyling/goal-engineering Reference implementation: GOAL.md schema, six canonical goal patterns, G0-G3 readiness scoring tool (Jun 2026) Dormant since 2026-06-30 (checked 2026-10-19: no commits since 2026-09-28; only release v1.1.0); repo description now targets Grok Build CLI /goal — low-yield
cobusgreyling/fleet-engineering github https://github.com/cobusgreyling/fleet-engineering Reference implementation: six fleet governance patterns, Fleet Economics cost attribution, F0-F3 maturity (Jun 2026)
omnigent-ai/omnigent github https://github.com/omnigent-ai/omnigent Meta-harness (4771★): 3-tier governance, harness-swap, cross-device session continuity (Jun 2026) 2026-10-19: ~100 commits in 3 weeks (mostly deps/tests/fixes) and weekly releases (v0.16.0 2026-09-29, v0.17.0 10-06, v0.18.0 10-13) — sweep releases first, not /commits; 4771★ not re-read
GitHub search — cross-model maker/checker github-search https://api.github.com/search/repositories?q=claude+codex+reviewer+loop&sort=updated Replaced low-yield acting_on query (0 results; that STATE.md gap is filled). Surfaces cross-model maker/checker harnesses (Claude implements / Codex reviews) — the Jun 2026 dominant theme 2026-10-19: 5 Tier-2 hits but 46 results, and ~2/3 of page 1 are 0-star single-author Claude+Codex reviewer-loop wrappers restating docs/04 (recency-dominated, not signal); a stars:>=5 floor is a possible refinement (untested). The unauthenticated API limit was hit after the first call, so dates are search-result pushed_at
getzep/graphiti github https://github.com/getzep/graphiti Temporal knowledge graph for agent state layer: invalidates stale facts, multi-modal search (vector+full-text+graph) (Jun 2026)
eugenelim/agent-ready-repo github https://github.com/eugenelim/agent-ready-repo Agent-ready repo spec: Surface vocabulary, adversarial reviewer checklists, verification modes, org learning stage (Jun 2026) 2026-10-19: description is stale — ~240 commits 2026-09-23..10-18; now carries a work-loop supervisor/cohort engine (ADR-0144), loop-cohort wave scheduling, structured review boundary, code-intelligence pack and a plan-evolution research record (docs/product/research/plan-evolution-experiments/FINDINGS.md). No GitHub Releases (pack versions appear in commit titles and docs/product/changelog.md): filter commits by paths under packs/core work-loop, docs/adr, docs/product/research; when the API is limited read github.com/.../commits/main with ?after= paging
JeiKeiLim/tenet github https://github.com/JeiKeiLim/tenet Structured loop critic framework: oracle problem, 3-tier document lifecycle, critic finding taxonomy (Jun 2026)
faisalishfaq2005/loopflow github https://github.com/faisalishfaq2005/loopflow Loop orchestration: cross-run memory persistence, gate feedback injection, debt-audit + docs-sync patterns (Jun 2026) Dormant since 2026-06-29 (checked 2026-10-19 via commits/main.atom: no commits since 2026-09-28; releases.atom empty)
uppifyagency/loop-kernel github https://github.com/uppifyagency/loop-kernel Provably-halting loop kernel: three exit-code stops (0/2/3), external unfakeable verifier, score= contract, LEDGER across compaction (Jun 2026) Dormant since 2026-06-28 (checked 2026-10-19 via commits/HEAD.atom; releases.atom and tags.atom empty)
affaan-m/ecc github https://github.com/affaan-m/ECC Flagship multi-harness "agent operator system" (224k★): eval-harness (pass@k/pass^k, 3 grader types), verification-loop, /loop-start, instinct learning; watch for focused loop-engineering skills (Jun 2026) 2026-10-19: canonical casing is ECC (lowercase still redirects); 224k★ not re-verified. ~875 commits 2026-09-26..10-15 (mostly fix/test/merge) and /commits/main.atom holds only ~20, so page the commits page's embedded JSON (react-app.embeddedData, commitsRefRoute.commitGroups) with ?after=+ in steps of 34; /releases.atom works and includes bodies. Most loop-relevant new work: a documented GateGuard hook-gate evaluation harness (docker/gateguard-effectiveness/, docs/gateguard/); v2.2.2/v2.2.3 were packaging only
krishddd/Strive_Engineering github https://github.com/krishddd/Strive_Engineering Provenance-bound verification: every finding cites a git SHA re-checked via git cat-file (loopguard), majority_vote blocks self-grading, L0-L3 autonomy ladder; isomorphic-perturbation verifier (Jun 2026)
Happenmass/omux github https://github.com/Happenmass/omux Renamed from Happenmass/Cliclaw 2026-09 (confirmed via GitHub API — the old repo ID no longer resolves; not a second repo from the same author, as this row previously read): cross-model maker/checker (Claude implements / Codex reviews) over tmux, auto-continue gate model, tmux-pane state scraping (hook-free), two-tier hybrid/SQLite memory; 95★ Quiet since 2026-07-14 (checked 2026-10-19 via commits/main.atom); latest Release/tag v3.0.1 (2026-06-26) — the 4.1.0 release commit has no Release or tag; 95★ not re-checked
firegnu/herdr-loop-lab github https://github.com/firegnu/herdr-loop-lab Three-layer (inner/fleet/epic) cross-model adversarial judge + mechanical gate; exit-code stop contracts (0/2/3), stateless worktree rounds, AC-N acceptance criteria (Jun 2026) Dormant: newest commit 2026-07-15, no releases or tags (checked 2026-10-19); deprioritise if it stays quiet
arXiv — harness/loop research rss https://export.arxiv.org/api/query?search_query=all:%22agent+harness%22+OR+%22loop+engineering%22+OR+%22self-improving+harness%22&sortBy=submittedDate&sortOrder=descending&max_results=30 arXiv Atom API; primary source for harness-engineering research (Self-Harness 2606.09498, AHE 2604.25850, HarnessX 2606.14249). Added Jul 2026 after 3 arXiv findings in one run Paging warning (2026-10-19): the harness topic yields ~150 papers per 3 weeks, so max_results=30 now spans only ~6 days (2026-10-12..10-16) and reading the stored URL alone misses anything older than a week — page with &start=30,60,90 (or raise max_results to 100) until last_run_date is reached (~100 hits in the 2026-09-28..10-19 window). Domain-specific harness papers (drug, video, medical, MOF) are omitted as applications
peterCheng123321/loop-engineering github https://github.com/peterCheng123321/loop-engineering Convergence layer over /loop, ralph-loop, Agent SDK loops; progress.md as DP "memo table" — cache solved steps, prune failed branches, survive compaction (Jul 2026)
Sungmin-Cho/deep-loop github https://github.com/Sungmin-Cho/deep-loop Control-plane/execution-plane split: kernel is sole authorized writer, skill agents read-only and write via kernel subcommands; content-hash-anchored state + append-only events across sessions (Jul 2026) Renamed from claude-deep-loop (old URL 301s, confirmed 2026-10-19). Releases v1.25.0 (2026-09-28) to v1.28.0 (2026-10-14)
shouryasrivastava/ctxcarry github https://github.com/shouryasrivastava/ctxcarry "Repo owns your context, not the agent": local-first .ctxcarry/ durable memory, worktree generators, evaluators-assume-broken, token-budgeted multi-tool handoff (Jul 2026). Default branch is master, not main — use /commits/master
the-open-engine/zeroshot github https://github.com/the-open-engine/zeroshot ~1.6k★; blind validation / information-asymmetry reviewers — validators see only outputs, never the maker's reasoning; anti-collusion reject-and-retry until all approve (Jul 2026) 2026-10-19: ~35 commits/week (mostly docs churn), v10.10.1 (2026-10-06) to v10.14.0 (2026-10-19) — read /releases; newer: customizable software-factory graphs, RuntimePlan per-node harness lanes, planner-led auto-research template (Oct 2026)
JasonxzWen/harness-hub github https://github.com/JasonxzWen/harness-hub Deep-read Jul 2026 (avg 3.5/5) — lock-versioned, state-separated distribution of skills/harness templates; a harness-marketplace angle distinct from typical skill package managers
edonadei/caliper github https://github.com/edonadei/caliper Deep-read Jul 2026 (avg 3.67/5) — local-first eval harness for Claude Code/Codex/Pi/Hermes skills; automatic with/without-skill baseline comparison isolates skill-attributable gains from base-model gains
explainx.ai html https://explainx.ai/blog/ Consistent loop/harness engineering coverage across 3+ runs (Loop Contract Model guide, context/prompt/loop/harness stack piece); overdue add per repeated "worth tracking" flags 2026-10-19: an RSS feed exists (https://explainx.ai/blog/rss.xml, also /rss.xml) but holds only the latest 60 items (~3 days at ~25 posts/day), so it cannot cover a 3-week window — enumerate slugs from the monthly sitemaps (https://explainx.ai/sitemaps/blog/blog-2026-10.xml, blog-2026-09.xml; ~1,500 posts, lastmod only) and read each article's JSON-LD datePublished; the /blog index lists ~20 posts and ?page=2 repeats it. Volume ~25-100 posts/day, mostly news rewrites with a small loop/harness share (~29 of ~400 keyword-matching slugs selected from title and lead) — daily runs can rely on the RSS
Daily Dose of Data Science rss https://www.dailydoseofds.com/rss/ Avi Chawla; 3+ high-engagement (100K-250K+ views) loop/harness/agentic-engineering explainers in one cycle, promoted via @akshay_pachaar (Jul 2026) Switched html -> rss 2026-10-19: valid RSS (30 items), full content:encoded for free posts; paid course parts truncate after the introduction (AI Agents course Parts 3-6 paywalled, so only the framing is verifiable). /feed redirects; /rss.xml and /sitemap.xml 404
hhamja github https://github.com/hhamja/claude-code-harness Renamed from hhamja/loop-harness (2026-09, confirmed via GitHub API repo-id redirect; old URL 301s), description unchanged. Prolific single actor flagged twice now (Jun/Jul 2026) with 5+ same-day loop-engineering/harness repos (loopkit, loopkit-b, my-loopkit, loop-harness, loop-engineering-architecture, claude-code-flywheel); this repo itself ships a cross-model verifier split and disk-based state — promoted to tracked after repeated "worth watching" flags Dormant since 2026-07-12 (HEAD 165a999, latest v0.5.0 2026-07-09; checked 2026-10-19 via main.atom/releases.atom) — downgrade from tracked if still silent next run
ruvnet github https://github.com/ruvnet Prolific actor, now 10+ relevant contributions: metaharness, ruflo, dream-machine, autogenous, openAVO, sparc, ruClip, RuVector, rvm, agentic-flow (multi-agent orchestration, 66 self-learning agents), open-claude-code (clean-room Claude Code CLI reimplementation) — an entire ecosystem of Loop Contract / self-improving-harness / fleet-infrastructure repos; track the profile for new repos, this list is not final (Sep 2026). 2026-09-21: profile "updated" badges on metaharness/sparc/rvm did not correspond to any main-branch commits after 2026-08-27/09-03 when checked directly — don't trust the badge alone, verify against /commits/main 2026-10-19: ruos dropped (not a repo; commits atom 404). No main commits on/after 2026-09-28 for sparc, autogenous, ruClip or openAVO (badge still unreliable). Active in window: dream-machine (nightly-routine engine, ADR-0011, software factory — most loop-relevant), ruflo (v3.57.0), metaharness (last main commit 2026-10-09), open-claude-code (daily upstream tracking); lower relevance: RuVector/agentic-flow/rvm, readme-proof, site-harness, rGi, APx, rultra, rudevolution. github.com/ruvnet.atom returned 504 twice: use ?tab=repositories&sort=updated plus per-repo commits atoms
ARC Prize Foundation rss https://arcprize.org/feed.xml Publishes independent harness-vs-model benchmark breakdowns (ARC-AGI-3) that repeatedly surface harness-engineering findings ahead of mainstream coverage — cited in docs/04, docs/24, docs/25 this run (Sep 2026) Switched html -> rss 2026-10-19: Atom feed advertised via link rel=alternate on the blog index (/blog/rss.xml 404). Newest post 2026-09-03, none in the window
coleam00/archon github https://github.com/coleam00/Archon 23,389★ — YAML-DAG harness builder (mixes deterministic and AI process nodes, isolated per-node git worktrees), by far the most externally validated harness-builder in this KB's corpus; near-daily merged PRs and semver releases every 1-2 weeks (Sep 2026) 2026-10-19: canonical slug is coleam00/Archon (lowercase redirects). Releases are the better signal (v0.11.x 2026-09-25, v0.12.0/.1 2026-10-17); the main feed shows only release-PR and Homebrew-formula commits; 23,389★ not re-verified (API rate-limited)
Hanako x https://x.com/hanakoxbt Published 2+ substantive loop-vs-graph explainer threads/videos this month (Aug 23 full course + Sep 5-6 "Loops vs Graphs" thread) with meaningful engagement (Sep 2026). 2026-09-21: crypto-promotion question settled — resolved, no action needed. A full profile scroll (13 days, 20+ posts) found zero $GRAPHRUN or other crypto-token promotion; the only crypto-adjacent element is that the linked paid course accepts USDT as a payment rail alongside Stripe, which is not token promotion. Separate, still-open fact-check flag: the account's Sep 9–19 post series opens each post with a quote falsely attributed to named public figures (Sam Altman with 4+ mutually exclusive wordings, a fake "Anthropic engineer", Karpathy), all funneling to the same paid course — never cite these attributed quotes as real; only the original loop/graph conceptual content is usable. Also flagged 2026-09-21: cross-promoted the same "Jev"/"Jev Engineering" product as @akshay_pachaar in the same window (Sep 15-20) with near-identical framing — treat as a single coordinated push, not two corroborating sources, if either is cited 2026-10-19: visible posts are dominated by REA (reverse-engineering agent) promotion, not loop engineering; the 14 Oct post again opens with a quote attributed to 'Anthropic engineers' — the same false-attribution pattern, never cite. Consider downgrading relevance if loop/graph content does not resume
Kirill x https://x.com/kirillk_web3 Published 2+ loop/graph-engineering guides this month (Aug 20 "Kimi K3 - From Loops to Graphs", Sep 6 "Loop and Graph Engineering — Full Guide") (Sep 2026)
Akshay Pachaar x @akshay_pachaar Co-founder @dailydoseofds_; loop engineering internals diagrams and practitioner breakdowns (Jun 2026). 2026-09-21: cross-promoted the same "Jev"/"Jev Engineering" non-LLM decision-engine product as @hanakoxbt in the same window (Sep 15-20) with near-identical framing — likely a coordinated sponsor push; treat as a single source, not two corroborating ones, if either is cited 2026-10-19 (logged-out): X search sits behind a login wall; visible posts are heavily Jev/TypeSafe promotion (10-19 'Anatomy of a Jev request', 10-16 'Jev is now multimodal') and the 17 Oct Switch video is a project promotion; profile order is not chronological (an 8 Jun article sits among 17-18 Oct posts), so it cannot serve as a timeline
ikangai rss https://www.ikangai.com/feed/ Independent blog running a coherent harness/loop/graph-engineering naming-progression thesis ("Graph Engineering Is Harness Engineering With a Diff" — cited in docs/21); worth periodic monitoring (Sep 2026) Switched html -> rss 2026-10-19: valid WordPress RSS 2.0 (10 items, newest 2026-10-17); ikangai.com redirects to www.ikangai.com. Only 2 posts in the 2026-09-28..10-19 window
Bruno Gonçalves / Data For Science rss https://data4sci.substack.com Primary source for original harness-engineering essays (OODA-loop/air-campaign framing) cited by explainx.ai's DAG-harness piece; most content is paywalled — only free previews are fetchable (Sep 2026) 2026-10-19: /feed is live, with full content:encoded for some posts and teasers for others; posts every ~2-4 weeks, 1 in the window (2026-10-09 vLLM); the agentic-harness series (05-27, 07-15, 08-12) pre-dates it. Low yield: mostly data-science, epidemiology and self-hosting topics
METR rss https://metr.org/feed.xml Source of transcript-tampering/tool-call-spoofing/reward-hacking research cited across multiple MindStudio posts this run (docs/17) — directly relevant to this KB's verification-loop and "a default that means success" material (Sep 2026) Switched html -> rss 2026-10-19: valid RSS (98 items, newest 2026-10-14, per-item pubDate; /feed, /rss.xml, /atom.xml 404). Window output is mostly AI-safety/control/evaluation (Senate testimony 09-30, Opus 5.5 pre-deployment evaluation 09-22, assessor-proposal posts 10-14); only the monitoring/untrusted-transcript pieces touch loop engineering: low-to-moderate yield
loop-js/loop.js github https://github.com/loop-js/loop.js 137★, not previously tracked. Goal/Round/read-only-Verify agent design with a concrete budget-ledger and crash-recovery mechanic — close match to this KB's Loop Contract (Sep 2026) Default branch is master (main.atom 404s). Checked 2026-10-19: newest commit 2026-08-30, newest release v0.2.0-beta (2026-07-10), nothing since 2026-09-28; low-activity, consider lowering priority
InjayTseng/loop-engineering-on-product github https://github.com/InjayTseng/loop-engineering-on-product Rare concrete before/after evidence (20 overnight rounds with a value gate vs. 112 iterations without) for why a value gate matters (Sep 2026)
huvii174/loop-engineering-plugin github https://github.com/huvii174/loop-engineering-plugin Six loop-engineering patterns with citable numeric thresholds (deterministic trigram-similarity circuit breaker, 95%-confidence design gate, tenth-man critic) — cited in docs/27 (Sep 2026) Releases/tags lag the repo (2026-10-19: latest tag v0.17.0 on 2026-09-22 vs CHANGELOG 0.19.1 on 2026-10-14) — sweep via commits/main.atom and CHANGELOG.md, not Releases
Kanevry/session-orchestrator github https://github.com/Kanevry/session-orchestrator Already the KB's Reference Implementations pick for orchestrator design (docs/32); actively released (v5.12.0, 2026-10-11; was v4.1.0 in Sep 2026) with a dedicated Loop Engineering course page — promoted to a tracked row so future releases are swept automatically rather than only surfacing via manual reading-list checks 2026-10-19: eight releases v5.5.0-v5.12.0 between 2026-10-02 and 10-11 (~1 per 1-3 days; v5.4.0 on 09-22) — high volume; releases.atom plus main.atom works (no REST API needed). Oct 2026 commit messages are largely German
milosCvetkovicDev/claude-code-monorepo github https://github.com/milosCvetkovicDev/claude-code-monorepo Real production config as a worked maker != checker example, with numbers (Sep 2026)
andrewcigan/vibe-dev-plugin github https://github.com/andrewcigan/vibe-dev-plugin Model-tiering-per-stage pattern (higher tier owns planning/critique/verification) with a physically write-denied verifier role — cited in docs/24 and docs/32 (Sep 2026)
dmlguq456/hearting github https://github.com/dmlguq456/hearting Sealed per-node model tiers + one harness-manifest.json projected across Claude Code/Codex/OpenCode's native surfaces — cited in docs/24 (Sep 2026) 2026-10-19: ~20 releases/day (v2.171 on 2026-10-01 to v3.36.10 on 10-19); the 10-entry releases.atom covers ~1 day, so a 3-week window needs the paginated /releases?page=N HTML (40 pages) — consider tracking only minor (vX.Y.0) releases. Release bodies are auto-generated PR titles only, so findings are title-level
firstintent/ccteam github https://github.com/firstintent/ccteam 425★. Cross-vendor, cross-machine agent-team coordination (Claude Code, Codex, Grok, Kimi, DeepSeek Harness, OpenCode, Pi) with a live cost ledger and per-harness quotas — cited in docs/23 (Sep 2026)
bluzername/loop-engineering-kit github https://github.com/bluzername/loop-engineering-kit Named slash-command loops operationalizing verification-as-engine + durable-memory ideas; good primary-source citations for Claude Code primitives docs (Sep 2026)
sdsrss/loop_eng github https://github.com/sdsrss/loop_eng Concrete Stop-hook implementation of a hard stopping condition with documented guard-degradation behavior — primary source for stopping-condition coverage (Sep 2026)
holny/Agent-Harness-Develop-Book github https://github.com/holny/Agent-Harness-Develop-Book Actively-maintained comparative source-code study across 7 agent harnesses, with a dedicated Loop Engineering chapter and an L1/L2/L3 governance pattern (Sep 2026) Last push 2026-09-24 (no push in the 2026-09-28..10-19 window; checked 2026-10-19)
loopx-project/loopx github https://github.com/loopx-project/loopx 5,916★, provider-neutral control plane implementing the KB's own Loop Contract concept at production scale; (Sep 2026). Merged 2026-10-19 with the former huangruiteng/loopx row — same repository (that URL 301s here; README credits Ruiteng Huang). Earlier note on it: 5,616★ (2026-09-05), local-first control plane above Claude Code/Codex/Cursor framed as "loop engineering" (quota-aware should-run gate, durable objectives+evidence logs — cited in docs/27 and docs/16) README now: 'open, local-first control plane for long-horizon agents' with an LHTB claim (LoopX 1.0.3 Heartbeat 0.4948 mean reward vs 0.4218 Plain Codex vs 0.4475 native Codex Goal, 46 tasks; authors caveat unequal budgets, not a causal estimate). Volume: 100+ commits/day, so commits.atom reaches back only ~4 hours; releases.atom is dominated by bot 'Desktop main ' previews, so read /releases tag pages, PyPI (pypi.org/pypi/loopx/json) and the project blog https://loopx-project.github.io/loopx/blog/; 5,916★ not re-measured
navels/neal github https://github.com/navels/neal Cross-vendor planner/coder/reviewer loop with a model-qualification harness producing a dated PASS/FAIL whitelist — concrete technique for maker/checker coverage (Sep 2026)
daonhan/ralph github https://github.com/daonhan/ralph Active implementation of the named "Ralph" loop pattern with a documented security trade-off (Docker socket bind-mount) — worth citing in unattended/headless-risk coverage (Sep 2026) 2026-10-19: feeds work without the API (default branch main). pnpm monorepo released by release-please (tags ralph-vX, ralph-core-vX, ralph-sandbox-vX), so releases.atom is mostly dependency-bump noise; last feed activity 2026-09-30
jamesst78/ship-loop-codex github https://github.com/jamesst78/ship-loop-codex States "the reviewer must never be the model that wrote the code" as a hard rule with a worked cross-model gate sequence; port of mstfash/skills — track together (Sep 2026)
mstfash/skills github https://github.com/mstfash/skills Upstream origin of the ship-loop pattern (see jamesst78/ship-loop-codex): plan-first, phase-gated loop with a definition-of-done gate (Sep 2026)
OutlineDriven/outline-driven-development github https://github.com/OutlineDriven/outline-driven-development 613-skill/28-plugin harness-agnostic distribution model — data point on plugin/skill ecosystem scale (Sep 2026) Mirror-sync maintenance (checked 2026-10-19): the only in-window commit is b0e8ce89 (2026-09-28, .devin skills sync), no loop-engineering content — low-yield
OrchestratorInc/agent-orchestrator github https://github.com/OrchestratorInc/agent-orchestrator 12,238★ desktop product implementing worktree-per-worker orchestration at scale — named large-scale adopter of the worktree pattern (Sep 2026) Moved from Untrivial-ai (old URL 301s, confirmed 2026-10-19); rebranded 'Orchestrator.inc' (v0.14.0, 2026-10-14). Stable releases every 1-3 days plus nightlies; the commit feed holds ~1.5 days, so read /releases (v0.13.2-v0.14.3 since 2026-09-28). Tier 3 only (UI/harness feature churn): low-yield, check monthly; 12,238★ not re-checked
Ilko Kacharov (kachar.dev) rss https://kachar.dev/rss.xml CTO/Co-founder Juma.ai. 4+ posts directly on-topic on loop/harness/verifier engineering (Sep 2026) Switched x -> rss 2026-10-19 (was tracked as https://kachar.dev/blog (X: @kachar136)): valid RSS (23 items, latest 2026-10-15), no Chrome needed. The Oct 2026 posts (drift, compound engineering, orders/questions) continue the on-topic rate
HarnessRouter/harnessrouter github https://github.com/HarnessRouter/harnessrouter Unified Harness Protocol (UHP) — cited in docs/24 since Sep 2026, promoted to a tracked row now that a second and third independent meta-harness product (openrig, orchestra) converged the same week (Sep 2026) 2026-10-19: 70 tagged releases v0.26.16 (2026-09-30) to v0.35.9, mostly github-actions[bot]; the feeds reach back only ~2 days, so page /releases?page=N (10 per page) and read protocol/CHANGELOG.md (best signal). ~60 patch releases unread; v0.32.0 (delete semantics) and v0.27.0 (Microsoft 365 plugin) seen but not reported
mvschwarz/openrig github https://github.com/mvschwarz/openrig 1,286★. "A harness wraps a model. A rig wraps your harnesses." YAML-defined multi-harness agent teams (Claude Code + Codex), tmux-backed persistent seats — cited in docs/24's meta-harness convergence note (Sep 2026) 2026-10-19: 10 detailed releases in the window (0.6.3 on 09-30 to 0.6.11 on 10-17) — releases.atom is the efficient path; 1,286★ not re-checked
DrSeedon/orchestra github https://github.com/DrSeedon/orchestra 15★. Independently arrived at this KB's own worktree-isolation + cross-model-review combination — small but conceptually clean, cited in docs/24 (Sep 2026) 2026-10-19: now 2.46.0 (10-15), ~20 commits on 10-19 alone; CHANGELOG.md is ~520 KB and bilingual (EN/RU) — read the v2.46.0 Highlights only. Its mutation-check merge gate (changed tests must fail against pre-change sources) is a verifier-design pattern worth a docs cross-check; 15★ not re-checked
cocodedk/loop-engineering github https://github.com/cocodedk/loop-engineering Fact-checked KB on Boris Cherny's "the loop," using the same confirmed-primary/confirmed-secondary/unverifiable sourcing discipline this repo uses — shared-origin caution applies (same caveat as ClaudeWarp), track for independent updates rather than cite as corroboration (Sep 2026) Last push 2026-09-24 (no push in the 2026-09-28..10-19 window; checked 2026-10-19)
ujconsulting/claudex-loop github https://github.com/ujconsulting/claudex-loop Cross-provider adversarial plan-hardening: Claude locks a plan, Codex attacks it read-only across bounded rounds, roles invert on build — cited in docs/04 (Sep 2026). Fork of chaseai-yt/claudex-loop, now tracked as its own row (github:https://github.com/chaseai-yt/claudex-loop). Standalone since commit 826e8f4 (2026-09-22) but cross-ports upstream issues by number (#33, #34, #28), so the two move together Default branch main; no GitHub Releases (releases.atom empty), so the commits atom is the only change signal (2026-10-19). Commit messages are partly German (docs/betrieb.md, README_DE.md)
chaseai-yt/claudex-loop github https://github.com/chaseai-yt/claudex-loop Upstream original of the tracked ujconsulting/claudex-loop fork (cross-provider adversarial plan-hardening). Added 2026-10-19: ~20 commits since 2026-09-06 with merged community PRs. No GitHub Releases (releases.atom empty), so sweep commits via /commits/main.atom; default branch main
Semih (@its_meseba) x https://x.com/its_meseba Publishing a numbered "Harness Engineering" series on X directly on loop/graph patterns for Claude Code — track for future installments (Sep 2026) 2026-10-19 (unverified): the only series visible on the logged-out profile is 'Part N in amazing repos' (Parts 46-49 posted 10-19, ~every 3 hours, repo spotlights); the 'Harness Engineering' series was not seen, so this rationale may be stale or older than the logged-out view reaches
Halldor Fannar linkedin https://www.linkedin.com/pulse/loop-engineering-halldor-fannar-acgzc/ Sr. Director at NVIDIA. Strongest single LinkedIn find to date: a real production loop (Rust clippy fixes, DGX Spark sandbox, state.json checkpoint/resume, 27 merges) — cited in docs/21 (Sep 2026) Re-read 2026-10-19 via curl: published 2026-07-29 (JSON-LD datePublished), unchanged, single static article already cited in docs/21 — no further yield; consider retiring from the sweep or keeping as a link-rot check. The NVIDIA title is not confirmed by the page body (it says only 'our Nemotron 3 Super model' and a disclaimer about 'my employer')

Type reference

Type Fetch strategy
x Two passes, both required. (1) Chrome → search from:<handle> (<tier1-tier2-query>) in X.com search. (2) Chrome → read the profile timeline at x.com/<handle> unfiltered, back to last_run_date or until the timeline stalls. Measured 20260906 on @bcherny: 7 posts / 4 posts / union 9 / overlap 2 — each pass is blind where the other sees. Report how far the timeline reached
rss WebFetch feed URL → parse <item> / <entry> → score against keyword tiers. Check that the feed's oldest item reaches last_run_date; if not, page it or use archive pages (2026-10-19: arXiv at max_results=30 spans ~6 days, The New Stack ~4, Simon Willison ~12, explainx ~3)
html WebFetch page URL → extract article links + snippets → score against keyword tiers
github WebFetch repo URL + /commits/main → score new commits since last_run_date; also check /releases for tagged releases. Unauthenticated api.github.com is limited to 60 calls/hour and was exhausted in the 2026-10-19 sweep — prefer /commits/<default-branch>.atom and /releases.atom (no API). The default branch is not always main (master for ctxcarry and loop-js/loop.js, where main.atom 404s); check it first
x-search Chrome → navigate to the search URL → scroll down ≥3 times to load 20+ posts → read all posts → score against keyword tiers
linkedin Chrome → navigate to the search URL → scroll down ≥3 times to load 20+ posts → use JavaScript to extract post text → score against keyword tiers; WebFetch any linked Pulse articles for Tier 1/2 matches. Content search only — see the ruled-out note below before adding a person-search row
github-search WebFetch the GitHub API URL (returns JSON) → parse items[].full_name, description, updated_at, stargazers_count → score against keyword tiers → for high-scoring repos WebFetch the raw README URL and summarise; check updated_at > last_run_date to flag new/recently-active repos

Ruled-out source strategies

Recorded so they are not re-proposed. A strategy that was measured and rejected is worth as much as one adopted — this KB has twice paid for un-annotated dead ends.

LinkedIn person-search — evaluated 2026-09-06, not adopted

The backlog asked for a linkedin person-search row (search/results/people/). A baseline was run before adding it, and the baseline argues against the row.

Query Kind Result
loop engineering Claude Code agent harness people 0 results
"loop engineering" people 10 shown, all matching on headline/skills text
"harness engineering" people 10 shown, same pattern
"loop engineering" OR "harness engineering" Claude Code (past month) content 3 substantive posts, 2 with Pulse articles

Two independent reasons it fails as a source row:

  1. It is network-biased, not topic-ranked. The same mutual connections recurred across two unrelated queries and results skewed to the operator's own city. The row would return whoever runs it — unreproducible across operators, and not a property of the field.
  2. Headline keywords do not predict published work. Deep-reading the strongest candidate (highest follower count, headline naming harness engineering) loaded 120 posts; 2 were on-topic. The rest was company marketing for an unrelated product.

Decisive comparison: none of the three authors the content search found appeared anywhere in the person-search results. They sit outside the operator's network, which is exactly where content search reaches and person search cannot.

Instead: the existing content row's keywords were broadened (above). Full evidence in plans/20260906_1259-c11-x-linkedin-baseline.md.

One finding was kept from the exercise: "loop engineering" has crossed into résumé vocabulary — it appears in LinkedIn headlines and skills sections across at least four countries, held by people who publish nothing on it. That is evidence about the term's diffusion, not a source lead. No individual is named: this repo is public and they are private individuals.


Adding new sources

Add a row if a general search surfaces someone who: - Published ≥ 2 pieces specifically on loop engineering, agentic AI, Claude Code, or MCP - Has meaningful audience engagement (quality over follower count)