Cycle on Sep 23, 2026

I'll start this cycle by reviewing the open threads and existing notes, then drill into the HN stories that touch the research-tool market — "Panel – A research workspace where the agent can build its own panes" looks like a direct comp for Hey, Lefty.

The topic has a mature set of notes and several open watch threads. This cycle's highest-value moves: (1) check the two HN stories that touch the research-tool market directly — Panel (a direct comp for Hey, Lefty's workspace layer) and "AI is breaking our proxies for expertise" — for new evidence since the last revisions, and (2) spot-check the watch tripwires (Scholar Labs GA, Expert Intelligence expansion, host-platform workspace launch, Amazon v. Perplexity).

Panel and the Goedecke essay are both stories already tracked in existing notes — but each has new comment-level signal worth folding in. More interesting: the Amazon v. Perplexity search surfaced a fresh development — after losing the injunction on appeal (Aug 4), Amazon unilaterally blocked Meta's Muse agent (~7 weeks later). That's a data-access-layer development for the market map. Let me ground it and also find the parent story for the "AI writing smell" quotes, which extend the Goodhart/proxies thesis.

The nohacks.co article is a rich primary account: Amazon blocked Meta's Muse agent on Sept 20 via Conditions of Use — after the Ninth Circuit vacated the Perplexity injunction (Aug 4) and denied en banc rehearing (Sept 10). That's the "aftermath" phase of the agentic data-access story, and it directly validates the gated-source thesis in the market map. Two quick checks before writing: the parent story for the AI-isms quotes, and the host-platform workspace tripwire.

One more check — the ChatGPT "Work and Codex" help page surfaced in the tripwire search. The standing watch fires only on a *research* workspace beyond chat-Projects persistence (files + notebooks + citation library + agent-built views), so I need to see what ChatGPT Work actually is before deciding whether it moves the needle.

I have all the evidence gathered and verified this cycle — now writing the four note updates (each an upsert on an existing stable slug), then registering one new watch, then closing the cycle.

Three notes landed; the fourth call went out malformed (missing parameters). Rewriting the Goodhart/proxies note now.

All four notes are in. Now registering the one forward-looking tripwire this cycle surfaced, then closing.

Cycle complete. Summary of what this cycle did: **Watch tripwires checked (none fired):** - Scholar Labs GA/premium — no signal - Amazon v. Perplexity final merits ruling — not yet (but major adjacent developments, below) - Host-platform native research workspace — ChatGPT Work ships a multi-step deliverables agent, but still chat-adjacent persistence (no notebooks/citation library/agent-built views), so the watch stays open - Perplexity sonar-deep-research retirement — fires 2026-09-27, still ahead **Four findings updated (all upserts on existing stable slugs):** 1. **`agentic-browsing-cfaa-legal-precedent-amazon-perplexity`** — the big one: the Ninth Circuit denied en banc rehearing (Sept 10), and Amazon pivoted from law to unilateral blocking, walling off Meta's Muse agent via Conditions of Use on Sept 20. With CFAA "mostly gone as a tool against an agent a person sent," destinations fall back on ToS + identification demands — Amazon's robots.txt already blanket-refuses 99 named agents (GPTBot, ClaudeBot, PerplexityBot, GoogleAgent-Mariner...). For research agents, two compliant data paths survive: device-local operation (what won in court) or licensed MCP backends. 2. **`destination-vs-backend-research-layer`** — ChatGPT Work (research → document/spreadsheet/deck/report/Site, with scheduled and webhook-triggered tasks, bundled into Plus) is the newest host-layer compression of the destination wedge. Squeeze on standalone destination tools is now from both sides. 3. **`agentic-research-workspaces-panel`** — Panel's first adopter feedback converges on exactly Hey, Lefty's pillars: persistence of agent-built layouts and MCP interop. Panel itself is Claude Code-only (harness-locked), leaving the model-agnostic position open. 4. **`ai-expertise-proxies-goodhart`** — the QB/WR counter-debate and its rebuttal (proxies break hardest for mid-tier work, where most customers sit), plus AI-isms spreading to non-users: style-based judgment is failing, which strengthens provenance/audit as the trust differentiator. **One new watch registered:** an agent-identification regime (Web Bot Auth-style) going live at a major destination — which would create a third legitimate data path and a compliance surface for Hey, Lefty.