Perplexity: search index as agent infrastructure

status: watch, opened 2026-09-09, last moved 2026-10-20. AI perplexity agents

The thread: Perplexity positioning its search index as the retrieval substrate for third-party agents — Search API inside Nous Research's Hermes Agent (open-weights), a 450B+ URL index with a stated path to a trillion URLs by EOY, and the Q2D-Web benchmark/leaderboard to set the eval terms for agentic RAG retrieval.
2026-09-09: opened from the Sep 9 ai digest (Arav RTs Sep 8 evening → Sep 9 4:21 PM ET).
Watch for: trillion-URL claim (ledger), Q2D-Web leaderboard entrants, whether other agent harnesses take the API, pricing.
2026-09-10: the thesis executes: the Perplexity API MCP server now supports OAuth (https://api.perplexity.ai/mcp, org sign-in, RT 11:08 PM ET Sep 9) — lowest-friction integration path yet; any MCP-speaking harness can add the index as a connector.
2026-09-10 (ops): the server exposes 4 tools — perplexity_search (raw ranked results, Search API, $0.005/request), perplexity_ask (fast preset, GPT-5.6 Luna, ~$0.003–0.01/answer), perplexity_research (high preset, GPT-5.6 Sol, ≤15 steps — the spend risk), perplexity_reason (medium preset, Luna, multi-hop). API-key bearer or OAuth 2.1+PKCE; local stdio server (@perplexity-ai/ mcp-server, MIT) for MCP clients without remote support. Rate limits by lifetime-spend tier (T0 1 QPS/50 RPM → T5 33 QPS/8k). Forcing function: Sonar Chat Completions deprecated Sep 27, 2026 — the MCP server is already on the Agent API.
2026-09-11: the Computer agent gets a revenue story. Arav: "Perplexity Computer makes you money" — underneath, Coinbase Developer Platform's own telemetry of agentic trading volume last week: Perplexity 26.8%, Claude 19.3%, Grok 15.3%, ChatGPT 1.1%, Claude Code 0.7%, CLI/other 36.8%. Single-source platform telemetry, "trading volume" undefined — but the first market-share read of agent economies, and Perplexity's agent leading it is a datapoint for this thread's thesis.
2026-09-14: the local turn — NVIDIA partnership + "Portable Computer" on Windows RTX PCs: harness/agents/models running locally, local files + connected apps without cloud, "unmetered local intelligence," frontier cloud only when needed (Arav, 11:28 AM ET). The index-as-agent-infrastructure thesis now has an edge variant with NVIDIA as the distribution channel; "unmetered" is marketing (local inference on the user's GPU costs Perplexity nothing). Same day: Computer adds @-coworker tagging in shared sessions — the human+AI-workspace direction. Watch: what "unmetered" actually meters, RTX-tier requirements, whether local mode still leans on the Search API backend.
2026-09-15: the agent-workforce claim, at infrastructure scale — CobbleDB, Perplexity's in-house DynamoDB replacement serving web content for search: "two engineers and hundreds of persistent Computer agents over two months… will save us up to a hundred million dollars yearly" (Arav, tweet ~12:30 PM ET; research being published). First concrete claim of core infra built by an agent fleet with two humans supervising; all self-reported, artifact is production search traffic, research forthcoming. Same day: Computer pre-installed on HP ZBook Ultra G3a (11:30 AM ET) — OEM distribution the day after the NVIDIA/RTX launch.
2026-09-17: spend controls ship — effort presets in Computer's model selector (Arav, ~12:08 PM ET): orchestrator model + reasoning depth combined per task, web now, more surfaces coming. The thread's spend-risk record (perplexity_research's ≤15-step risk) gets its vendor-side control.
2026-09-24/25 (distribution day): AMD partnership — Portable Computer on Ryzen AI Max processors ("agentic PC" framing; OEM layer now NVIDIA RTX + AMD); DocSend connector live ("Deal Room" use case); Photon Search API claims (Rust-based, sub-250ms p95); Fast Search — free for all Nous Portal tiers, now the default search in Hermes Agent. The index-as-agent- infrastructure thesis keeps executing on distribution; trillion- URL claim (ledger) still pending.
2026-09-28: three compounding moves. (1) Marquee partner in NVIDIA's Open Agent Safety Platform (OpenShell + Sentry, 100+ partners) — Arav pledges "we intend to open source all of it"; Perplexity's month-long SPACE red-team (9 models, root access) is the receipt to watch (openai-agent-incidents has the containment layer). (2) Agent API grows Profiles, Skills, and managed connectors — configurable reusable agents. (3) Arav's research call (continual learning, multi-agent parallel workers, synthetic-data evals — "we make enough money to fund new research"), amplified by LeCun — the research-in-public posture extends from infra to science.
2026-09-29: two agentic-claim leaps from Computer (both vendor demos, not audited): (1) multi-agentic trading on Coinbase — Computer running Fable 5.1, GPT-6 Astra, and Grok 4.7 "in tandem" on an agentic portfolio with automatic research/rebalancing (sources: public earnings, CB Insights, on-chain) — the agent-economics discourse (Apollo bank-run thought experiment, Sep 28) gets its first big-vendor claim on real market rails; (2) an end-to-end browser-based open-world game built by Computer in Standard Mode (Opus 5.5) for ~$2,000 in credits — rented GPUs, headless Blender, procedural design, assets, audio (vice.pplx.app). Watch: third-party reproduction, what the Coinbase loop does with real funds.
2026-10-04 (platform layer): Arav — "You can build custom vertical AI apps inside Computer": demo GeoGuessr-LLM built inside Computer (Perplexity SDK web search, local place lookups, browser control, Cesium for 3D/satellite; ~3:20 PM ET). Follows the Vera silicon announcement: Computer is being pitched as a vertical-app platform, not just a consumer assistant. Demo-grade, vendor-run, unaudited.
2026-10-12 (~8:09 PM ET): Perplexity Projects — every Project is a Git repo: "a Git forge for agents, without the GitHub bloat, since no human has to approve anything. Each agent gets its own short-lived [credentials]" (Aravind Srinivas) + the filesystem build-out RT. Development of the substrate layer: version control purpose-built for machine actors. Watch: the per-agent credential pattern vs the record's agent-incident threads (Hugging Face breach, DseWiki, Gemini test breach), adoption beyond Perplexity's surface. Same window: a one-person company spending ~$1M/year on Perplexity (Molly O'Shea interview) — the consumer-enterprise boundary blurring.
2026-10-15 (ai-social window 2, ~3:33 PM ET — https://x.com/AravSrinivas/status/2110816382099325030): SPACE goes open-source — WSJ's Reed Albergotti (RT'd by Arav Srinivas): "Perplexity is open sourcing its sandbox, called SPACE, and has partners like Akamai, Coreweave and Crusoe in the big push for AI security," topped with the neocloud thesis: "Neoclouds can't just lease GPUs. They need to own the full agent inference platform: open-weight models (SGLang), fine-tuned models (Miles, River), and secure sandboxes (SPACE)." Development on the Sep 28 entries — the "we intend to open source all of it" promise now in execution, and the sandbox that ran the pending 9-model root red-team moving onto open operator paper; on the containment thread this is the first open-source sandbox layer beside NVIDIA's closed Open Agent Safety Platform. Same window: Nemotron 3.5 Super VL ("NVIDIA's Nemotron 3.5 Super VL is now available in Perplexity for Pro and Max subscribers, and in Agent API," ~2:59 PM ET — https://x.com/AravSrinivas/status/2110807863123935732) — another third-party model on the distribution surface beside Opus 5.5 and GPT-6 Sol. Watch: the SPACE repo/license landing, the still-owed red-team writeup (what broke with root access, promised Sep 28), partner integrations, adoption beyond Perplexity's surface.
2026-10-15 (~1:03 PM ET artifact landing — https://x.com/AravSrinivas/status/2110880073394139580): the checkable item from the same day's open-sourcing announcement resolved — Perplexity's own post (via Arav RT): "We're open sourcing SPACE, the sandbox platform that runs every task in Perplexity Computer. SPACE gives each agent task its own Firecracker microVM and controls credentials and outbound traffic from outside the VM. Available now under Apache 2.0," repository at github.com/perplexityai/SPACE. Same window, distribution layer: Perplexity API now available inside Lovable apps, "real-time answers from the web, with citations" (~6:14 PM ET — https://x.com/AravSrinivas/status/2110856743148519578). Watch: the still-owed Sep 28 red-team writeup, Akamai/Coreweave/Crusoe integrations, adoption beyond Perplexity's surface.
2026-10-14 (window 2, connector + tier layer): (1) Ultrafast on Computer (~1:58 PM ET — https://x.com/AravSrinivas/status/2110430065708376566) — covered in the GPT-6 Astra rollout entry. (2) Polymarket connector (~2:11 PM ET — https://x.com/AravSrinivas/status/2110433423894651279): "You can now tag @Polymarket inside Perplexity Computer for market and real time trends research tasks" — Perplexity Finance pulling real-time data from Polymarket's API; market-implied probabilities, trading volume, liquidity "across 250,000+ active prediction markets." The thread's connector ledger's first prediction-market-class integration — a market-data rail beside the agent-trading record (Coinbase telemetry Sep 11, multi-agentic Coinbase trading Sep 29). Vendor-shipped claims only: no independent receipts, no credit/pricing mechanics behind the "credit drain warning" in the window. The same day's business-metrics post (Arav, ~3:59 PM ET — https://x.com/AravSrinivas/status/2110460403159540041) goes to tick-2026-10-14-ai-agent-economy-adoption-reads.
2026-10-19 (~2:47 PM ET — https://x.com/AravSrinivas/status/2112254336160108664): SPACE gets its first named partner integration — Crusoe. Crusoe: "Agents need secure sandboxes as much as they need GPUs. We're partnering with @perplexity_ai to bring SPACE to our Managed AI portfolio on Crusoe Cloud." Arav Srinivas: "Crusoe is extremely well positioned to become a full-stack cloud: power, compute, and services running on top. Perplexity's open source sandbox platform SPACE will become part of Crusoe's Managed AI products. Look forward to more collaborations with them!" Development of the Oct 15 open-sourcing entries: the first of the named partners (Akamai, Coreweave, Crusoe) formalizes, and the neocloud thesis ("Neoclouds can't just lease GPUs… they need to own the full agent inference platform") moves from tweet to product placement. Announcement-grade: no availability date, technical detail, or pricing in capture; the still-owed Sep 28 SPACE red-team writeup remains unwritten. Watch: a SPACE-on-Crusoe live date; Akamai/Coreweave follow-ons.
2026-10-20 (~12:34 PM ET — https://x.com/AravSrinivas/status/2112583333825950207): the local turn reaches macOS — Perplexity Computer "can now run fully local on your Mac" with "no cloud based inference," model and agent both on-device, working with native Mac apps and local files, and tasks run "without sending sensitive data to the cloud"; the product account states it is "available now in the Perplexity Mac app." Continuation of the Sep 14 NVIDIA/RTX local turn; the "available now" is a vendor-status claim with no independent check in the window. Watch added: which Mac hardware runs the full stack, and whether local mode still leans on the Search API backend.