Autonomous Research Tools: Market Map Digest
TL;DR
Perplexity went on a three-part offensive this week — Automations (always-on agents), Computer for Enterprise (a 20-model "meta-router" pitch), and Portable Computer (official local execution) — which together close off most of the whitespace an orchestration-layer product thought it owned (perplexity-deep-research-consumer-agent). Meanwhile, a fourth wave of discussion around Sean Goedecke's expertise-proxies essay sharpened the cultural argument for audit-grade research tools: when output can't be trusted at face value, verifiable process becomes the scarce good (ai-expertise-proxies-goodhart
). The squeeze on Hey, Lefty's positioning is real from both directions — but the two forces point at the same wedge.
Perplexity is now selling the orchestrator pitch, not just the answer
The destination layer's model-agnostic pitch is no longer a startup's differentiator — Perplexity is selling it to enterprises as the whole point. Forkast's analysis of the Ask 2026 unveil names the strategy directly:
"Perplexity isn't trying to build the next frontier model. Instead, they are positioning themselves as a meta-router—an orchestration engine that sits on top of the existing AI ecosystem." — Perplexity's bet on the 'meta-router' model — Forkast
The enterprise post claims it "routes work across 20 models, using the right model for each task," with 400+ app connectors and reported assignments like Claude Opus 4.6 for reasoning and Gemini for deep parallel research (perplexity-deep-research-consumer-agent). For the market map, this means "don't pick winners, route across models" — a plausible core value prop for a new entrant — is now the incumbent's B2B headline, and demand is evidently there: Forkast reports "over 100 enterprise customers demanded access over a single weekend following their consumer launch" (perplexity-deep-research-consumer-agent
).
What to watch: Whether Perplexity publishes a public list price for Computer for Enterprise — Forkast reports pricing is "usage-based... with an organization-wide credit pool" with no public number, the last missing consumer-side price point for the destination layer.
Always-on agents and local execution arrived in the same week
Two capability thresholds crossed at once. On the proactive side, Perplexity's Automations launch brings scheduled and event-triggered long-running agents — "Unlike other agents, Automations don't start from scratch every time" — with triggers from Slack, Gmail, Outlook, Linear, or GitHub, and credits spent "only when Computer runs an assignment" (perplexity-deep-research-consumer-agent). That converges with OpenAI's dots on the same pattern: memory-carrying, execution-metered agents that watch rather than wait.
On the local side, Portable Computer is a full harness, not a thin client — "the model, orchestrator, scheduler, and sandbox on your own device," with local MCP server support that "cloud Computer does not do" (perplexity-deep-research-consumer-agent). But the fine print defines the remaining moat: it runs only 27B-class local models (PPLX 27B, Qwen 27B) on NVIDIA hardware with 24 GB of VRAM or more, requires a Pro/Max subscription, and drops memory — "Brain is a cloud feature... not available in local mode." A true local-first, model-agnostic orchestrator — any model the user brings, no subscription gate, full local autonomy — still occupies distinct ground, but "runs on your device, private" is now table stakes (perplexity-deep-research-consumer-agent
).
What to watch: Whether local mode reaches frontier-class models, Apple Silicon, or persistent memory — any of the three would narrow the local-first differentiation window fast.
The expertise-proxy debate is writing Hey, Lefty's positioning copy for it
The cultural argument for audit-grade research tools got sharper this week, as the fourth discussion wave of Sean Goedecke's essay landed on Hacker News. The essay's Goodhart mechanism: "Puzzles were a useful, impossible-to-game measure for mathematical progress. But now that AI companies can game that measure (by solving them in a way that's inaccessible to humans), the whole point of those puzzles disappears" (ai-expertise-proxies-goodhart).
The new wave's sharpest exchange is precisely the research-tool buyer's dilemma. One commenter defends the proxies with a football analogy — AI is the wide receiver, the human the quarterback — and the rebuttal turns it:
"If AI is the wide receiver, it's perceived currently by many to be an absolutely elite top-tier receiver. Especially at lower leagues, QBs who have these clearly elite WRs are discounted and considered more-or-less replaceable all the time." — Comment by BoxFour
That's the risk in reverse for research tools: when the tool is an elite receiver, the human's contribution stops being legible — and gets undervalued. Goedecke's software-engineering extension lands the same way: a meaty GitHub project "now projects like that are worthless, because everyone just assumes they're vibe-coded" (ai-expertise-proxies-goodhart). For positioning, this is the argument that tools making the human's contribution and the machine's chain of evidence visible are selling the repair of a broken proxy, not just answers — and that unbranded AI research output will trade at a discount as provenance becomes the scarce good.
What to watch: Whether the "legible process" wedge gets a commercial champion — a research tool marketing auditability as the headline feature rather than a compliance footnote.
What surprised us
- The local-first turn came from the cloud incumbent, not the privacy startups. Portable Computer runs the entire orchestrator, scheduler, task queue, local search index, and sandbox on-device — a far deeper local implementation than expected from the destination layer (perplexity-deep-research-consumer-agent
). The counterintuitive gift: it escalates frontier reasoning back to Perplexity's cloud and gates itself behind a subscription, so the "no gatekeeper, any model" niche survives.
- The vendor ROI numbers are doing heavy lifting unaudited. Forkast flags Perplexity's "internal study of 16,000 queries that allegedly resulted in $1.6 million in labor cost savings and 3.25 years of work completed in four weeks" as "vendor-reported metrics and have not been externally audited" (perplexity-deep-research-consumer-agent
- The Goodhart debate resolved into a genuine stalemate, not a consensus. The Thurston/Naur camp ("AI generated proofs short-circuit human understanding") versus the Tao camp ("Terence Tao uses AI heavily and has been writing extensively about how useful it is") are still talking past each other (ai-expertise-proxies-goodhart
) — which is exactly the unresolved tension a provenance-first tool can sell into.
Open threads worth a vote
- [watch] Perplexity publishes Computer for Enterprise list pricing — the missing data point for the destination layer's pricing segmentation; steer follow-up on the topic page.
- [watch] Portable Computer local mode reaches frontier-class or expands beyond NVIDIA — the trigger conditions that would compress Hey, Lefty's local-first window; steer follow-up on the topic page.