Autonomous Research Tools: Market Map Digest — The Search Stack Gets Given Away, and Agents Start Reviewing the Reviewers
TL;DR
Two moves this cycle redraw adjacent edges of the map. Perplexity — the destination-layer leader — open-sourced a full multimodal retrieval stack under MIT, commoditizing what many research-tool startups would call their core technology (retrieval-substrate-open-source). Meanwhile a Show HN launch called agent.reviews opened a fight over a layer nobody had mapped yet: who controls what tools agents recommend (agent-native-tool-discovery
). Both land on Hey, Lefty's positioning questions — one as a gift, one as a warning.
Perplexity is commoditizing the retrieval layer it sits on top of
The destination layer's leader has stopped treating search infrastructure as a moat and started giving it away. On October 7, Perplexity released pplx-embed-v2-late — late-interaction embedding models at 0.6B and 9B scales, multimodal over text, images, and rendered PDFs with no OCR pipeline, on Hugging Face under MIT with commercial use allowed (retrieval-substrate-open-source). The blog names the deployment modes explicitly, and one of them is practically a product spec for a privacy-positioned research agent:
"Local–cloud retrieval: encode private local documents or live queries with the 0.6B model on-device, and compare or merge those representations with results from a 9B index hosted in the cloud." — Multimodal embeddings beyond a single vector — Perplexity blog
This is the third Perplexity open-infra release in three weeks — Photon (Sept 25), CobbleDB (Sept 16), plus an open-weights Decisions API model — which reads as a deliberate strategy: monetize the destination and the hosted API, commoditize everything underneath (retrieval-substrate-open-source). For the market map, the "core technology" column for the search layer now has an answer: it's commoditizing fast, under permissive licenses. For Hey, Lefty specifically, the 0.6B model (~1.2 GB of weights, laptop-class) means SOTA multimodal retrieval on-device without paying a retrieval vendor or leaking corpus contents to one. One caveat carries over from last cycle's Goodhart discussion: the benchmark numbers (92.4% on MADQA, 64.0% on BrowseComp+) are self-reported, with a full technical report only "later this year."
What to watch: The hosted API endpoint is "planned but not live" — when it ships with published pricing, self-hosting economics finally get a list price to benchmark against (retrieval-substrate-open-source).
A discovery chokepoint is forming over what agents choose — and it's already gaming itself
When agents do the choosing, vendors will optimize for the agent the way they optimized for Google — and the infrastructure for that fight appeared this cycle. "Agent.reviews – Where AI agents read and write reviews on tools" hit Show HN at 72 points and 49 comments, with reviews written by users' agents rather than humans; the founder is associated with Armature, a YC-backed company whose pitch is to "Be the tool Claude Code chooses" (agent-native-tool-discovery). The founder's own framing names the power problem outright:
"Claude Code for example picks Anthropic's code review tool in 80% of the cases. At some point this will probably face antitrust considerations but for the time 3rd-party vendors need to survive. We don't want a world where only labs survive, do we?" — screm, in the Show HN thread
The thread's peer review surfaced the structural failure modes, and they rhyme with last cycle's Goodhart essay: grade inflation ("cryptography got 4.6/5 stars… YAML also got 4.6/5 stars. I'm thinking the agents are grading on a curve" — themgt), a prompt-injection surface, and the expected monetization endgame — "pay-to-rank-higher as the next move… Adwords for tooling!!" (agent-native-tool-discovery). Commenters also flagged Stack Overflow's agents.stackoverflow.com as prior art, meaning this is a category forming, not a one-off launch. For the map, whoever aggregates agent-written experience data sits on the research-tool market's next discovery chokepoint — and Hey, Lefty is on both sides of it: its own recommendability inside harnesses is a distribution channel to manage, and agent-written review corpora are a candidate input source with exactly the provenance problems above.
What to watch: Any sign of paid placement on an agent-reputation platform, or a major harness shipping an official tool-recommendation channel — either hardens discovery as a purchasable chokepoint.
What surprised us
- The destination leader is arming its own competitors. Perplexity's MIT-licensed retrieval stack is usable by any orchestrator — including ones positioned against Perplexity's own product — while its only monetization (the hosted API) doesn't exist yet (retrieval-substrate-open-source
). That's a land-grab on infrastructure, not revenue.
- The Goodhart essay's prediction showed up in a launch thread within weeks. Agent-written star ratings inflating toward 4.6 across unrelated categories, with the founder conceding "agents never put 1-star rating so we may need to normalize the scale" (agent-native-tool-discovery
) — output proxies collapsing in real time, now in tool discovery rather than expertise.
- Perplexity's decontamination stance reads as a shot at the benchmark-industrial complex: "we exclude any datasets associated with benchmarks that we evaluate on… trustworthy numbers are more useful to us" (retrieval-substrate-open-source
- The HN discussion of the embedding release was nearly empty (2 points, 0 comments as of Oct 11, per the HN story) while the agent.reviews thread ran 49 comments deep — the audience cares about who chooses tools, not how retrieval works.
Open threads worth a vote
- [watch] pplx-embed family hits the Perplexity API Platform rate card — steer what counts as firing: hosted endpoint live with published pricing.
- [watch] Agent-native discovery layer monetizes or gets a harness-default channel — steer between paid placement, harness defaults, and vendors openly budgeting for Agent Experience as the trigger.