Cycle on Sep 20, 2026

I'll start by reviewing the topic's open threads and full note set, then drill into the HN stories that touch this market (the Panel research-workspace launch and the "AI is breaking our proxies for expertise" piece look directly relevant).

The note set is rich and was updated as recently as yesterday. Two HN stories I'd flag (Panel, the Goedecke essay) are already covered by existing notes (`agentic-research-workspaces-panel`, `ai-expertise-proxies-goodhart`). So this cycle's value is in checking the standing watches and scanning for anything that landed in the last 24–72 hours across the key players. Let me scan for fresh developments.

Scholar Labs shows no GA/premium news (that watch stays open). But there are genuinely new signals: Perplexity shipped a local "Portable Computer" for Windows with scheduled deep research, a "Premium Data Sources" tier (FactSet, PitchBook, CB Insights…), a frontier-model training deal with Crusoe, and a new backend player called Parallel selling "Deep Research" as web infrastructure. Let me read the primary sources.

Strong haul — three substantive developments. Before writing, let me check the one new HN launch that touches the workspace/orchestrator layer (Skillsync, 64 pts) and run a quick check on the Kimi international-expansion watch.

Skillsync isn't in the cached scan — let me read its discussion directly, since session-portability across agents is adjacent to the orchestrator-layer thesis.

Skillsync's discussion is about coding-agent session portability — a useful secondary signal on the lock-in question, but not a core research-tool story. One more diligence check: the standing watch on host platforms (OpenAI/Anthropic/Google) shipping a native research workspace.

The Google "notebooks in Gemini" rollout could be a material update to the existing NotebookLM-rebrand finding — let me read it before deciding.

The TechRepublic fetch failed. Let me confirm the OpenAI "Library" feature from its primary help-center page instead — it's the relevant signal for the host-workspace watch.

You're right — I have the evidence gathered; now I'll persist it. Three writes: a new note on Parallel (new market entrant), and revisions to the Perplexity and bifurcation notes, all grounded in sources fetched this cycle.

All three findings are persisted. No watches fired this cycle (Scholar Labs stays waitlist-gated, OpenAI's Library is files-only so the host-workspace watch holds, Kimi shows no international launch, and the Perplexity sonar retirement lands 9/27). The cycle is complete.

Cycle complete. Here's what this round surfaced for the market map: **New finding — Parallel (parallel.ai), the backend pure play.** The strongest exemplar yet of research-as-infrastructure: Task/Search/Extract/Monitor/FindAll APIs on a proprietary index, priced per request ($1/1,000 for search; "pay for answers, not tokens"), with the Basis framework attaching calibrated confidence scores + citations + reasoning traces to every output. It publishes head-to-head benchmarks vs Exa/Tavily, counts Harvey/Pfizer/Notion/Manus as customers, and — notably — pays publishers by Shapley value for their marginal contribution to agent work. For Hey, Lefty it's both a potential supplier (MCP server, free tier) and the evidence-standards bar to meet. **Perplexity's triple move (revised note).** In one week it went vertical across every layer: **Portable Computer** for Windows (model + agent harness + orchestrator + scheduler entirely on-device, credit-free local work, permission-gated cloud escalation), a **premium data tier** (bundled Wiley/PitchBook/CB Insights/Statista/Midpage plus ~30 BYO connectors incl. FactSet, Guidepoint, Crunchbase), and the **Crusoe deal** to train its own frontier models. Its boilerplate now calls Perplexity Computer "a model-agnostic orchestrator of AI agents" — converging on the orchestrator layer from the consumer side. **Bifurcation thesis updated.** The backend layer is thickening into real vendors while destinations absorb infrastructure — and at the session layer, Skillsync (Launch HN, 64 pts) is making agent sessions portable across harnesses ("You should not get locked into one agent simply because you built up a conversation history"). Net: value compresses everywhere except Hey, Lefty's layer, but the bar rises — orchestration must now match backend economics and Basis-grade confidence scoring. **Watches checked, none fired:** Scholar Labs still waitlist-gated; OpenAI's new Library is files-only (no notebooks/citation library), so the host-workspace watch holds; no Kimi international launch; Perplexity's sonar-deep-research retirement lands 9/27.