Cycle on Oct 11, 2026

I'll start by reviewing the topic's open threads and existing notes to see where this cycle should pick up.

The topic is deep into a mature market-map research effort: 40+ notes, many watches. Today is Oct 11, 2026; the last cycle updated notes on Oct 10. The HN map's only story ("How Protein Took over the World" — paywalled FT, 0 comments) is off-brief noise for an AI-research-tools market map, so I'll focus this cycle on fresh developments since Oct 10 that touch the market map's layers. Let me sweep for recent news and HN discussion.

The Perplexity sweep surfaced two fresh items this week: the open-sourced `pplx-embed-v2-late` embedding models (Oct 7) and the Rings AI connector for Perplexity Computer (Oct 6). The "deep research" search mostly returned stale April/2025 material. Let me check HN for Perplexity/Sacra threads and find the official embedding announcement.

Strong lead: Perplexity published an official blog post on `pplx-embed-v2-late` (Oct 7) — MIT-licensed multimodal late-interaction embeddings, which speaks directly to the brief's "core technology" and the local-first layer of the market map. Also an HN hit: "Agent.reviews — where AI agents read and write reviews on tools" (72 pts, 49 comments, Oct 7), a possible new agent-native discovery layer. Let me read both.

The Perplexity blog post is fully read — it's a substantive release (MIT-licensed, 100% local inference mode, local–cloud hybrid retrieval). The Agent.reviews story isn't in the cached scan, so let me read its comments directly.

The Agent.reviews thread is rich — founder-linked to Armature ("Be the tool Claude Code chooses"), with HN calling it "SEO for agents" and predicting pay-to-rank. That's a genuinely new subject for the market map: an agent-native discovery/reputation layer. Before writing, let me confirm the MIT license claim from a fetched source and date the Snap–Perplexity deal that surfaced in the first sweep.

I have two findings worth persisting from this cycle's research. Writing both notes now.

Both notes are saved. Now registering the two forward-looking watches this cycle surfaced.

Cycle complete. Summary of what this cycle did: **Findings written (2):** 1. **[[retrieval-substrate-open-source]]** — Perplexity's Oct 7 release of `pplx-embed-v2-late` (MIT-licensed, multimodal late-interaction embeddings, 0.6B/9B with a shared embedding space) puts SOTA retrieval on laptops ("100% local inference") and enables hybrid local–cloud merging — the enabling infrastructure for the local-first layer of the market map, and evidence the search layer's "core technology" is commoditizing under permissive licenses while Perplexity monetizes the destination and upcoming hosted API. 2. **[[agent-native-tool-discovery]]** — the agent.reviews Show HN thread (72 pts) exposed a forming agent-native discovery/reputation layer, tied to Armature's "Be the tool Claude Code chooses" pitch, with HN surfacing its structural problems: prompt-injection surface, agent grade inflation, cold-start incentives, and an expected pay-to-rank endgame. Relevant to how research tools (including Hey, Lefty) get chosen when agents do the choosing. **Watches registered (2):** pplx-embed hitting the Perplexity API Platform rate card (one_shot), and the agent-native discovery layer monetizing or gaining a harness-default channel (ongoing). **Skipped, deliberately:** the HN map's only story (paywalled FT protein piece, 0 comments) is off-brief; the Snap–Perplexity $400M deal's mutual end is stale Q1 2026 news; the decider-model-as-grader question thread stays open (no new evidence this cycle).

The research cycle is complete. Two new findings were persisted to the topic's living set — the open-source retrieval substrate (Perplexity's pplx-embed-v2-late and what it enables for local-first research agents) and the emerging agent-native discovery layer (agent.reviews/Armature) — plus two watches registered so future cycles check whether the retrieval substrate gets list prices and whether agent-mediated tool discovery starts monetizing.