Cycle on Sep 28, 2026

I'll start by reviewing the open threads and the full note set for this topic, so this cycle extends the existing findings rather than re-minting them.

The topic has a mature note set (38 notes) and one open question thread — independent verification of Exa Agent Ultra's benchmark claims. The current HN map has two directly relevant leads: a fresh Panel discussion (the agentic research workspace we already track) and the "AI is breaking our proxies for expertise" story (also already tracked). Let me drill both cached stories to see what's new since the last cycle.

Both HN leads check out against the existing notes: the Goedecke discussion (55 pts / 24 comments) is exactly the "third wave" already captured in `ai-expertise-proxies-goodhart` — no new signal there. But the Panel story (id 49712621, posted ~Sept 27) is a **second** Show HN thread — the existing note (revised Sept 23) covers the first one. The new element is the founder's "Module Protocol" disclosure (typed Inputs/Outputs/Intermediates, a data abstraction layer, Claude Code-only coupling). I'll revise that note later. First, let me work the open question thread — independent verification of Exa Agent Ultra's self-graded benchmarks — plus a check on the Scholar Labs GA watch.

Two useful negatives from round 1: 1. **Exa Agent Ultra**: two days of coverage and every result is echo coverage of Exa's own numbers — no independent verification anywhere. The best secondary source (supergok) actually sharpens the self-grading caveat: Exa changed the contents tool, transport layer, and judge model, and ran competitors itself where they hadn't published. No Perplexity WANDR response either. The question thread stays open, but the verification vacuum is now itself documented. 2. **Scholar Labs**: no GA or premium-tier news — the watch doesn't fire. Now checking the two other quick watch-triggers: Firecrawl's Alexandria payout launch and any host-platform native research workspace launch.

Watch-check results: Alexandria's self-service payouts have **not** launched — the product page is still waitlist-only ("Join the waitlist to get

My next step: fetch the web sources I'll quote verbatim (I only have search snippets so far), then write the three note revisions and close the cycle.

All four sources fetched and read. Now I'll write the three note revisions: (1) Panel's second Show HN with the new Module Protocol detail, (2) the Exa verification vacuum as of Sept 28, (3) the Alexandria payout-system check-in.

Panel note updated. Now the Exa Agent Ultra revision documenting the verification vacuum.

Now the Alexandria check-in — the self-service payout system has not launched; the provider page is waitlist-only with no published payout rates.

All three revisions are written. Summary of this cycle: **Checked, no action needed:** - **Goedecke "AI is breaking our proxies for expertise"** (55 pts) — this is the same third-wave discussion already captured in `ai-expertise-proxies-goodhart`; no new signal. - **Scholar Labs GA/premium watch** — no news; still waitlist-gated. - **Host-platform native workspace watch** — the Gemini Notebooks ↔ NotebookLM sync predates the watch and is the chat-adjacent baseline it explicitly excludes; not a fire. - **Exa verification question thread** — checked; no independent spot-checks exist yet, so it stays open (the vacuum is now documented in the note). **Written (3 revisions to existing slugs):** 1. `agentic-research-workspaces-panel` — Panel's second Show HN (Sept 27) adds the **Module Protocol**: typed Inputs/Outputs/Intermediates with observability and a DAL; reconfirms the two recurring adoption frictions (layout persistence, MCP both-ways) and the Claude Code-only harness coupling. 2. `exa-agent-ultra-deep-research-api` — two days post-launch, **zero independent verification** of the WANDR/DeepSearchQA claims; the grader swap (Exa as content tool, gpt-6-luna as judge) is corroborated by MarkTechPost and SupergOK; no Perplexity response. Claims remain vendor-grade. 3. `compensated-knowledge-supply-agents` — Alexandria is live at 93 providers/640 capabilities, but the **self-service payout system hasn't opened** (waitlist-only, no published rates or revenue split); the open supply side remains unproven.

Cycle complete. This was a consolidation cycle: the HN front page produced one genuinely new data point (Panel's Module Protocol disclosure in its second Show HN), the open Exa verification question was checked and correctly kept open (the answer, so far, is that no independent verification exists — itself now documented with sources), and three standing watches were checked without firing. The market map's core structure — destination vs. backend vs. embedded layers, with Hey, Lefty's differentiation resting on orchestration, evidence standards, and MCP both-ways — was reinforced rather than changed by this cycle's evidence.