Autonomous Research Tools: Market Map Digest — Cycle 2
TL;DR
The bifurcation from last cycle's digest is no longer a hypothesis — it now has a scaled pure play on each side. Parallel is selling deep research as per-request backend infrastructure with audit-grade provenance, while Perplexity has become the most vertically integrated destination player in the market, spanning data bundles, local orchestration, and its own frontier models. For Hey, Lefty, both moves sharpen the same conclusion: everything below the orchestrator layer is commoditizing fast, and the bar for the layer above it keeps rising.
The backend layer found its pure play — and it pays publishers
The destination-vs-backend question now has a definitive test case. Parallel (parallel.ai) sells no destination product at all — its entire suite (Search, Extract, Responses, Task, FindAll, Monitor) is API-only, priced per request: "Pay for answers, not tokens," with Search starting "as low as $1 per 1,000 requests" (parallel.ai). Its published benchmarks put Parallel Advanced at 74% on BrowseComp at $399 per 1,000 requests, versus Tavily frontier at 66%/$935 and Exa frontier at 70%/$971 — deep research as a priced, comparable commodity.
The provenance story is the most developed in the market, and it's quotable in full:
"You can trust an answer in production and audit it afterward, tracing any claim back to the page it came from." — Parallel
Its Basis framework attaches "calibrated confidence scores, citations, source excerpts, and reasoning traces" to every output, and the underlying Parallel Web Index spans "hundreds of billions of pages" including licensed paywalled content. The most consequential mechanism is Index by Parallel, which pays content owners by Shapley value — "its marginal contribution to the work an agent performs at the moment of inference" (parallel-web-systems-research-backend). That's the first market mechanism aligning publisher incentives with agentic research, as opposed to the litigation path. The customer list — Harvey, Pfizer, Manus, Notion, Dropbox, Hex — reads like a who's-who of products that need research plumbing but won't build it.
What to watch: Whether Shapley-value payouts get adopted by a second backend vendor — if they do, publisher economics become a standard feature of the backend layer rather than one vendor's differentiator.
Perplexity is trying to be every layer at once
Perplexity answered the vertical-integration question with a yes-to-all. In mid-September it shipped Portable Computer for Windows, which runs "the model, agent harness, orchestrator, and scheduler entirely on the Windows device" — with on-device work consuming no cloud credits and cloud escalation permission-gated ("When a task needs to send something from the device to the cloud, Portable first asks for the user's permission") (Perplexity blog). It requires an RTX GPU with at least 24GB of VRAM, so this is a prosumer/enterprise wedge, not a mass-market one.
On data, it now bundles licensed premium sources — Wiley, PitchBook Essentials, CB Insights, Statista, and Midpage's 13.8M U.S. court opinions — plus a bring-your-own-entitlement connector catalog spanning FactSet, Morningstar, Guidepoint, Crunchbase, and dozens more (Perplexity Help Center). And on models, the Crusoe partnership (September 15) has it training frontier models on dedicated GB300 clusters. Its own boilerplate now calls Perplexity Computer "a model-agnostic orchestrator of AI agents across tools, files, chips, devices, and more" (Crusoe announcement).
This matters for the map because Perplexity is converging on the orchestrator/workspace layer from the consumer side — the same seat Hey, Lefty wants. Its blind spots define the wedge: evidence-grade audit trails and research-specific workflows are not what it emphasizes (perplexity-deep-research-consumer-agent).
What to watch: Whether the premium-data bundle (included sources with daily query limits) becomes the standard consumer packaging for licensed content — it's a materially different model from Elicit's per-seat pricing or Parallel's per-request economics.
The hosts keep inching toward the workspace — and sessions are becoming portable
The standing tripwire from last cycle — a chat host shipping a full native research workspace — still has not fired, but OpenAI moved closer. Its new Library layer gives ChatGPT persistent, searchable, shareable files across conversations, with connected Google Drive/Box/Dropbox/SharePoint sources and plan-tiered storage from 500MB (Free) to 100GB (Pro) (OpenAI Help Center). It's files-only — no notebooks, citation library, or agent-built views — but persistence inside the host app is exactly the capability that erodes a standalone workspace's reason to exist.
Meanwhile, a counterforce to lock-in appeared at the session layer. Skillsync (YC W26, Launch HN) unifies agent session transcripts across harnesses into one schema. Co-founder Narsagna's framing:
"You should not get locked into one agent simply because you built up a conversation history + memory with that agent." — Narsagna on Launch HN
The payoff is trajectory-level continuity: "You can continue a teammate's session on your own machine, or pick up a claude.ai conversation in Claude Code or Codex. The new agent gets the transcript: all the decisions in the trajectory, not a summary of it" (destination-vs-backend-research-layer). If conversation history stops being a moat, workspaces must win on something else — which is, again, evidence standards and orchestration quality.
What to watch: Whether Library grows a notebook or citation layer within a quarter — that's the moment the host-workspace tripwire actually fires.
What surprised us
- The backend pure play arrived before most people noticed the category existed. Parallel's per-request pricing (74% BrowseComp at $399/1k requests) makes "deep research" a line item an engineer adds to an invoice, not a product decision (parallel.ai). Anyone still pitching "AI deep research" as the product is already behind.
- Perplexity's data catalog quietly includes expert transcripts and court opinions. Guidepoint expert transcripts plus Midpage's 13.8M-case citator bundled into a consumer subscription is institutional-grade data at Pro-tier prices (Perplexity Help Center) — a sharper data play than Kimi's, and aimed at individuals rather than institutions.
- Publisher economics may get solved by markets, not lawsuits. Parallel's Shapley-value payouts give content owners a reason to license into agentic search rather than litigate it — if the mechanism holds, the "AI vs. publishers" conflict narrative ages badly.
- Session portability is an anti-lock-in force nobody had on the map. Skillsync's premise — transcripts as portable assets, not vendor property — attacks the persistence advantage that destination products (including Elicit's multiplayer workspace) quietly rely on.
Open threads worth a vote
- [watch] Host platform ships a full native research workspace — Library moves the needle but remains files-only; the tripwire is still untriggered (destination-vs-backend-research-layer
) destination-vs-backend-research-layer
- [watch] Does Parallel's Shapley-value publisher payout get copied, and does Perplexity's included-premium-sources bundle become the consumer standard? Both determine where the data layer's margin settles (perplexity-deep-research-consumer-agent
) destination-vs-backend-research-layer