Week 13 briefing — 2026-07-05

config/cloud-network-allowlist.md

@@ -0,0 +1,96 @@
+# Cloud run — network allowlist for the Learning Agent
+
+> **Why this file exists:** On 2026-07-05 a full sweep confirmed that 38/40 source endpoints
+> fail from the cloud (Claude Code on the web) environment with `CONNECT tunnel failed,
+> response 403` — the **environment's egress network policy** blocks them. This is NOT
+> YouTube/Substack rate-limiting a datacenter IP (the rss2json fallback proxy was blocked
+> too, which a site-side block can't explain). The previous runs' "datacenter IP block"
+> diagnosis was wrong. No fetch trick fixes this — only the environment settings can.
+
+## How to fix (one-time, ~2 minutes)
+
+In **claude.ai/code → Environments → (the environment this repo's scheduled task runs in)
+→ Network access**, either:
+
+- **Option A (simplest):** set network access to **Full / unrestricted**. This repo's agent
+  only reads public feeds and writes to Notion/GitHub, so this is low-risk. **Recommended.**
+- **Option B (tighter):** keep the allowlist mode and add the domains below.
+
+Docs: https://code.claude.com/docs/en/claude-code-on-the-web
+
+## Domains to allowlist (Option B)
+
+Verified blocked on 2026-07-05; each maps to a source in `config/sources.md`.
+
+### Critical — primary content sources
+```
+youtube.com            # all 13 YouTube channel RSS feeds (www.youtube.com/feeds/videos.xml)
+substack.com           # Substack redirect/edge for all custom-domain feeds below
+substackcdn.com        # Substack CDN (feed assets/redirects)
+oneusefulthing.org     # Ethan Mollick (Tier 1)
+simonwillison.net      # Simon Willison (Tier 1)
+lennysnewsletter.com   # Lenny Rachitsky (Tier 1)
+smashingmagazine.com   # Vitaly Friedman (Tier 1)
+maggieappleton.com     # Maggie Appleton (Tier 1)
+medium.com             # Google PAIR feed (Tier 1) + Aakash mirror
+shapeof.ai             # Emily Campbell (Tier 1, manual check)
+wattenberger.com       # Amelia Wattenberger (Tier 1, manual check)
+```
+
+### Tier 2 feeds
+```
+blog.google
+openai.com
+deepmind.google
+huyenchip.com
+interconnects.ai
+jack-clark.net
+magazine.sebastianraschka.com
+cameronrwolfe.substack.com
+latent.space
+stratechery.com
+newsletter.pragmaticengineer.com
+creatoreconomy.so
+exponentialview.co
+jakobnielsenphd.substack.com
+lukew.com
+newsletter.uxdesign.cc
+bradfrost.com
+peterme.com
+garymarcus.substack.com
+langchain.com
+news.aakashg.com
+feeds.simplecast.com   # Hard Fork podcast feed
+every.to               # Dan Shipper / AI & I
+pair.withgoogle.com
+aixdesign.co
+normaltech.ai
+```
+
+### Useful extras
+```
+api.rss2json.com       # fallback feed proxy (only needed if a site later rate-limits the runner)
+nngroup.com            # recurring discovered-pick source
+arxiv.org              # recurring discovered-pick source
+```
+
+### Already reachable (no action)
+```
+microsoft.com / www.microsoft.com   # MS Research feed — worked all along
+raw.githubusercontent.com          # Anthropic engineering community RSS mirror (stale, but reachable)
+```
+
+### Known special case
+```
+anthropic.com   # bypasses the session proxy entirely (NO_PROXY) and Anthropic's edge
+                # blocks datacenter fetches with a real HTTP 403. Allowlisting may not
+                # help; the agent should keep using WebSearch for anthropic.com articles.
+```
+
+## After changing the setting
+
+The next scheduled run re-tests automatically: STEP 2's fetching discipline starts with a
+one-call preflight and reports `policy-blocked` vs `OK` per lane in Source Health. If the
+briefing still says `policy-blocked` after you've updated the environment, the change
+didn't apply to the environment the task actually runs in — check which environment the
+scheduled task uses.

config/previously-recommended.md

@@ -66,7 +66,6 @@ https://www.youtube.com/watch?v=D4XTefP3Lsc
 https://www.deloitte.com/us/en/insights/topics/talent/human-capital-trends/2026/human-ai-interaction-design.html
 https://newsletter.uxdesign.cc/p/on-rational-thinking-taking-action
 
-<<<<<<< Updated upstream
 # Week 6 — 2026-04-20
 https://maggieappleton.com/zero-alignment/
 https://www.youtube.com/watch?v=QVJcdfkRpH8
@@ -83,7 +82,7 @@ https://www.microsoft.com/en-us/research/blog/red-teaming-a-network-of-agents-un
 https://www.youtube.com/watch?v=mCO-D3pkviM
 https://simonwillison.net/2025/Jun/27/context-engineering/
 https://uxdesign.cc/the-web-trained-ai-to-deceive-now-designers-have-to-untrain-it-420f9d3fe78e
-=======
+
 # Week 6 — 2026-04-11 (Test run — retroactively logged)
 https://smashingmagazine.com/2026/04/identifying-necessary-transparency-moments-agentic-ai-part1/
 https://jakobnielsenphd.substack.com/p/workflow-redesign
@@ -110,4 +109,37 @@ https://engineering.salesforce.com/grounding-enterprise-ai-with-live-web-retriev
 https://www.smashingmagazine.com/2025/07/design-patterns-ai-interfaces/
 https://mbrenndoerfer.com/writing/rag-architecture-retriever-generator-design-patterns
 https://www.youtube.com/watch?v=5ABcYSWPVr8
->>>>>>> Stashed changes
+
+# Backfill — prior Week 13 briefings (were posted to Notion but never logged here; added 2026-07-05 to fix dedup drift)
+# Week 13 — 2026-06-14
+https://www.anthropic.com/engineering/built-multi-agent-research-system
+https://aakashgupta.medium.com/building-multi-agent-systems-will-be-a-must-have-pm-skill-in-2026-30e52fd732cb
+https://www.lennysnewsletter.com/p/we-replaced-our-sales-team-with-20-ai-agents
+https://www.latent.space/p/andon
+https://www.microsoft.com/en-us/research/blog/magenticlite-magenticbrain-fara1-5-an-agentic-experience-optimized-for-small-models/
+https://arxiv.org/abs/2510.24937
+https://arxiv.org/abs/2603.02050
+
+# Week 13 — 2026-06-21
+https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents
+https://learn.microsoft.com/en-us/microsoft-copilot-studio/guidance/multi-agent-patterns
+https://arxiv.org/abs/2512.11979
+https://agentic-design.ai/patterns/ui-ux-patterns
+https://www.news.aakashg.com/p/practical-ai-agents-pms
+https://www.oneusefulthing.org/p/what-it-feels-like-to-work-with-mythos
+
+# Week 13 — 2026-06-28
+https://www.anthropic.com/engineering/harness-design-long-running-apps
+https://jakobnielsenphd.substack.com/p/no-more-ui
+https://www.lennysnewsletter.com/p/make-product-management-fun-again
+https://www.nngroup.com/articles/state-of-ux-2026/
+https://www.news.aakashg.com/p/ai-agents-pms
+https://www.oneusefulthing.org/p/co-existence-and-the-end-of-co-intelligence
+
+# Week 13 — 2026-07-05 (fourth pass)
+https://www.oneusefulthing.org/p/the-twilight-of-the-chatbots
+https://www.lennysnewsletter.com/p/not-all-ai-agents-are-created-equal
+https://www.anthropic.com/engineering/how-we-contain-claude
+https://www.nngroup.com/articles/context-architecture/
+https://jakobnielsenphd.substack.com/p/2026-predictions-halfway
+https://simonwillison.net/2026/Jul/4/better-models-worse-tools/

config/sources.md

@@ -56,7 +56,6 @@
 | 47 | Harrison Chase | 2 | Agents | Blog | blog.langchain.dev | https://www.langchain.com/blog/rss | Surface architecture posts. Skip product announcements. Feed updated 2026-04-18 — old URL `blog.langchain.com/rss` permanent-redirects to this one. |
 | 48 | Nadia Piet / AIxDESIGN | 2 | Design x AI | Blog | aixdesign.co | | Manual check monthly. Surface toolkit updates. |
 | 49 | Design Better | 2 | Design Leadership | YouTube | youtube.com/@designbetterpod | https://www.youtube.com/feeds/videos.xml?channel_id=UCMVvDE1tbXNoOpoUaMk6iUw | Casual viewing. Surface design leadership + AI design interviews. Skip episodes on non-AI topics unless guest is exceptional. |
-<<<<<<< Updated upstream
 | 50 | Aakash Gupta — Product Growth | 2 | Product | Substack | news.aakashg.com | https://www.news.aakashg.com/feed | Surface AI-for-PM posts: evals for PMs, AI product playbooks, agent product strategy. Practitioner PM voice — Lenny-tier for product. Skip pure career/interview-prep posts. (Feed verified live 2026-06.) |
 | 51 | Hard Fork (NYT) | 2 | Strategy / Culture | Podcast | nytimes.com/column/hard-fork | https://feeds.simplecast.com/l2i9YnTd | Kevin Roose + Casey Newton. Audio-only (no YouTube) — judge from episode title + description. Surface AI policy, product launches, agent/Copilot, and AI-culture segments. Skip non-AI tech-news episodes. (Feed verified live 2026-06.) |
 | 52 | Every — "AI & I" (Dan Shipper) | 2 | Building / Product | Podcast + Essays | every.to/podcast · every.to/@danshipper | No RSS — manual | How real people (founders, writers, designers) actually use AI day-to-day. No clean RSS on Every's platform. Manual/discovery: WebFetch every.to/podcast + the AI & I YouTube playlist (PLuMcoKK9mKgHtW_o9h5sGO2vXrffKHwJL), and lean on web search for "Dan Shipper" / "AI & I". Closest in spirit to Lenny. |
@@ -81,7 +80,6 @@ Curated handles to check as a **secondary, best-effort** signal (see the X step
 | 2 | @aakashg0 | Aakash Gupta | AI for PMs, evals, product playbooks. |
 | 2 | @danshipper | Dan Shipper | Every / AI & I — how people actually use AI. |
 | 2 | @hwchase17 | Harrison Chase | LangChain; agent architecture. |
-=======
 
 ---
 
@@ -105,5 +103,3 @@ Proposed new Tier 1/2 voices to fill the product + design gap. Katherine reviews
 2. Confirm posted content in last 90 days (active signal)
 3. Read 2-3 recent posts to confirm lane + Tier
 4. Add to live table with number, write Agent Notes, done.
-
->>>>>>> Stashed changes

logs/run-2026-07-05.md

@@ -0,0 +1,44 @@
+# Run Log — 2026-07-05
+Week: 13 (Multi-agent orchestration — design lens)
+Topic: Week 13 · Multi-agent orchestration (design lens)
+Sources checked: ~25 (13 YouTube channels attempted, Tier-1 blog/Substack feeds attempted, 12 web searches; X skipped)
+Items collected: ~18 candidates surfaced via web search
+Items after hard filters: 6 (excluded all URLs from the three prior Week 13 briefings 06-14/06-21/06-28, the Weekly Checklist, and previously-recommended.md)
+Items after topic gate: 6
+Must Read: 2 (score range: 94–96) — Mollick "Twilight of the chatbots", Lenny "Not all AI agents are created equal"
+Recommended: 3 (score range: 76–90) — Anthropic "How we contain Claude", NN/g "Context Architecture", Jakob Nielsen "2026 AI & UX mid-year reality check"
+Wild Card: yes (trigger: Contrarian) — Simon Willison "Better Models: Worse Tools"
+Discovered: 0 (all six are known Tier-1/2 voices; no new voice cleared the bar this run)
+Dropped: several low-trust SEO/listicle "agentic design patterns 2026" pages; additional Simon Willison July posts (AI Compass, DSPy Datasette, June newsletter) as lower topic-fit; Aakash Gupta posts (avoided — prior briefings already leaned heavily on Aakash for the 2nd Product slot)
+
+## Idempotency (STEP 1d)
+Guard did NOT trigger. Three "Weekly Picks — Week 13" pages already exist (created 2026-06-14, 06-21, 06-28). Most recent is 7 days old — outside the 3-day window — so this run proceeded. New page created: https://app.notion.com/p/39482e5a5fa38177b2cff67f21d299dd
+
+## Diversity rule
+Satisfied: 2 Product Lane (Mollick, Lenny), 2 Design Lane (NN/g, Nielsen), 2 pure-engineer-audience (Anthropic containment, Willison) — at the ≤2 engineer cap. No shortfall to note.
+
+## Errors/Notes
+- YouTube RSS + rss2json proxy: wholesale HTTP 403 (datacenter-IP block), same as the three prior runs. No fresh-video pick.
+- Blog/Substack feeds (direct AND rss2json proxy): 403. Anthropic community RSS mirror reachable but stale (Nov 2025). Article bodies also 403 on direct fetch from this IP. Entire run rebuilt via web search + prior knowledge of the named sources — matching how the three prior Week 13 runs operated.
+- Dedup drift fixed: previously-recommended.md only went to Week 7 and was missing the ~18 URLs from the three prior Week 13 briefings. Backfilled all of them plus this run's 6.
+- prompts/weekly-briefing.md and config/*.md still contain unresolved git merge-conflict markers (Updated upstream vs Stashed changes). Honored both sides this run (STEP 0 sync + idempotency guard from one side; audience framing + Audience Lanes/diversity rule from the other). Flagging for cleanup.
+- STRONG SIGNAL: fourth consecutive Week 13 briefing. Katherine has left every Week 13 item unchecked across four runs while the checklist algorithm keeps returning Week 13 as "current." Fresh Tier-1 multi-agent-orchestration content is exhausted. Recommend advancing the checklist to Week 14 (Agent-Computer Interface + sub-agents). Surfaced prominently in the Notion page callout, the @mention, and the push notification.
+
+## Deliverables
+- Notion page: https://app.notion.com/p/39482e5a5fa38177b2cff67f21d299dd
+- @mention comment to catphy0825@gmail.com (Katherine): posted (id 39482e5a-5fa3-8186-a801-001d01a8c5b4)
+- Google Calendar event: intentionally NOT created (per task instruction; Notion notification is sufficient)
+- PushNotification: sent
+Duration: ~single run
+
+## Addendum — 2026-07-05 (post-run diagnosis correction)
+The "datacenter-IP block" diagnosis in this log (and in the three prior Week 13 runs) was WRONG.
+A curl sweep of all 40 source endpoints shows `CONNECT tunnel failed, response 403` — the
+**cloud environment's egress network policy** blocks 38/40 hosts (only microsoft.com and
+raw.githubusercontent.com are allowlisted). The requests never reached YouTube/Substack;
+that's also why the rss2json fallback "failed" — it was blocked too, not overwhelmed.
+Fix requires a settings change by Katherine (environment network access), documented in
+config/cloud-network-allowlist.md. weekly-briefing.md STEP 2 now includes a preflight that
+distinguishes `policy-blocked` from real rate-limits. Merge-conflict markers in
+prompts/weekly-briefing.md, config/sources.md, and config/previously-recommended.md were
+also resolved (both sides kept).

prompts/weekly-briefing.md

@@ -6,7 +6,6 @@ You are Katherine's Learning Partner Agent. Your job is to find the highest-qual
 
 ---
 
-<<<<<<< Updated upstream
 ## STEP 0 — Sync repo state (1 tool call)
 
 This task runs on more than one machine (Katherine's laptop + a cloud fallback). Before reading anything, pull the latest committed state so the dedup log is current:
@@ -16,15 +15,16 @@ git pull --rebase
 ```
 
 If this fails (no network, or not a git checkout), continue anyway — the run still works, it just may not see the very latest dedup entries.
-=======
+
+---
+
 ## Who this is for
 
 **Audience:** Katherine is a future PM + designer learning AI to *apply*, not to build from scratch. She is not an engineer and will not write production AI code. She prefers content that helps her make product decisions, design human-AI interactions, evaluate tradeoffs, and speak fluently with AI engineers — not content that teaches her to be one.
 
 **When scoring, ask "who is this written FOR?"** If the answer is "engineers building this system," the item can still appear but should not dominate the top of the briefing. Prefer content written for practitioners who *use* AI to build products and experiences.
 
 This audience framing is enforced downstream by the **Audience Check** (pre-scoring), the **Audience Lane bonus** (scoring), and the **diversity rule** (final-7 selection). All three must agree.
->>>>>>> Stashed changes
 
 ---
 
@@ -72,18 +72,29 @@ Run two complementary scans. Budget your tool calls carefully:
 - **YouTube engagement check**: ~3-5 WebFetch calls (top candidates only)
 - **Notion write**: 1 call
 
-### Fetching discipline — beating 403 rate-limits and IP blocks
-Feeds (especially YouTube, Substack, Medium) return **HTTP 403 when the requester's IP is rate-limited or blocked** — the signature is "the first few succeed, the rest 403," or (on a datacenter/cloud IP) *wholesale* 403s. The feeds are not dead; the IP is blocked. Escalate **per feed**:
+### Fetching discipline — 403s come in two flavors; diagnose FIRST, then escalate
+
+**PREFLIGHT (1 tool call, do this before any feed batch):** fetch one YouTube feed, e.g.
+`https://www.youtube.com/feeds/videos.xml?channel_id=UCYO_jab_esuFRV4b17AJtAw` (3Blue1Brown).
+- If it succeeds → normal run, go to Flavor B discipline below for any later 403s.
+- If it fails, run `curl -sS -o /dev/null -w "%{http_code}" <same URL>` in Bash. If curl reports **`CONNECT tunnel failed, response 403`** (or exit code 56), you're in **Flavor A**.
+
+**Flavor A — environment egress-policy block (cloud runs).** The session's network policy is refusing the tunnel; the request never reaches YouTube/Substack at all, and the rss2json proxy is equally unreachable — *no fetch path or proxy trick will work*. (Confirmed 2026-07-05: 38/40 source endpoints blocked this way; only microsoft.com and raw.githubusercontent.com were allowlisted.) Do NOT burn tool budget retrying every feed or the proxy:
+1. Spot-check at most 2 more hosts (one Substack, api.rss2json.com) to confirm it's wholesale.
+2. Fall back to targeted WebSearch for the entire run (search still works — it's a harness-side tool).
+3. In Source Health, report the lanes as **`policy-blocked`** (not "Failed"/"IP block") and link `config/cloud-network-allowlist.md` — the fix is Katherine adding the domains (or enabling full network access) in the environment's settings, and no run-side change can substitute.
+4. Note it under Errors/Notes in the run log.
 
+**Flavor B — target-side rate-limit / datacenter-IP block.** The signature is *mixed* results: some fetches return real content or real HTTP status pages, then YouTube/Substack/Medium start returning 403/429. The feeds are not dead; the requester IP is throttled. Escalate **per feed**:
 1. **Fetch the feed URL directly** (fast path — works from a residential IP / laptop). Don't burst: fetch in **small sequential batches (3–4 at a time)**, not all at once.
 2. **On a 403/429, refetch the same feed through the free rss2json proxy:**
    ```
    https://api.rss2json.com/v1/api.json?rss_url={URL-ENCODED feed URL}
    ```
-   URL-encode the feed URL — YouTube feeds contain `?channel_id=` and must be encoded (e.g. `https%3A%2F%2Fwww.youtube.com%2Ffeeds%2Fvideos.xml%3Fchannel_id%3D...`). The proxy fetches from its own un-blocked servers and returns the feed as JSON (`status:"ok"` + `items[]`). **This is what makes cloud/datacenter runs work** — YouTube/Substack/Medium block datacenter IPs directly but not the proxy. (Verified working for YouTube, Substack, and Medium.)
+   URL-encode the feed URL — YouTube feeds contain `?channel_id=` and must be encoded (e.g. `https%3A%2F%2Fwww.youtube.com%2Ffeeds%2Fvideos.xml%3Fchannel_id%3D...`). The proxy fetches from its own un-blocked servers and returns the feed as JSON (`status:"ok"` + `items[]`).
 3. **Only if the proxy also fails**, fall back to a targeted web search for that source.
 
-Never silently degrade to web-search-only. In Source Health, say which path worked, e.g. `YouTube: OK (via rss2json proxy)`.
+Never silently degrade to web-search-only, and never mislabel Flavor A as an IP block. In Source Health, say which path worked per lane, e.g. `YouTube: OK (direct)`, `Blogs/RSS: OK (via rss2json proxy)`, or `YouTube: policy-blocked (see config/cloud-network-allowlist.md)`.
 
 ### 2a. FRESH SCAN — What's new this week (last 7 days)
 
@@ -467,12 +478,10 @@ Before creating the Notion page, verify:
 - [ ] Checklist at bottom has ALL of: tier emoji, title, estimated time, full URL for EVERY item
 - [ ] Total estimated reading time in header is the sum of all item times
 - [ ] Hard cap of 7 items total is respected
-<<<<<<< Updated upstream
 - [ ] Idempotency guard checked (STEP 1d) — not re-posting a week that already ran
 - [ ] PushNotification sent (STEP 7) and state committed + pushed (STEP 8)
-=======
 - [ ] **Audience Check recorded for every item** (engineers / PMs / designers / mixed / general)
 - [ ] **Diversity rule satisfied**: ≥2 Product Lane, ≥2 Design Lane, ≤2 pure-engineer-audience items
 - [ ] If diversity rule couldn't be satisfied by the pool, the shortfall is noted in the run log under Errors/Notes
 - [ ] No orphan references to the old "+5 Design Boost" — the rule is now three Audience Lanes (+10/+10/+5)
->>>>>>> Stashed changes
+- [ ] Source Health distinguishes `policy-blocked` (environment egress) from target-side failures — see Fetching discipline