Autonomous Research Tools: Market Map Digest — Cycle 3
TL;DR
The compression on Hey, Lefty's layer is now coming from three directions simultaneously: Perplexity is shipping the orchestrator down to the desktop OS and the silicon, session portability is dissolving the conversation-history moat that destination products rely on, and the agent-builds-its-own-workspace mechanic went from novel to replicated in a single week. The scarce ground is narrowing to typed research workflows, evidence standards, and orchestration over research-grade data sources — everything else is commoditizing in public.
Perplexity is now an orchestrator company — and it's buying the distribution to prove it
The consumer agent platform has stopped describing itself as a search engine. Its own boilerplate in the Crusoe announcement now calls Perplexity Computer "a model-agnostic orchestrator of AI agents across tools, files, chips, devices, and more" (Crusoe announcement). Last cycle's digest flagged the Portable Computer for Windows launch as a prosumer wedge; this cycle extends it into an OEM beachhead — the HP deal pre-loads the Perplexity app in the Windows taskbar starting with the ZBook Ultra G3a (Perplexity blog).
The pricing architecture is the quietly radical part. On-device work is credit-free, and cloud escalation is permission-gated:
"When people run work on device, it doesn't consume Computer credits, making long-running workflows more practical." — Perplexity blog
That decouples long-running scheduled agents from metered usage entirely — a fundamentally different economic model from per-request backends like Parallel or per-seat tools like Elicit. Meanwhile the Crusoe deal covers the full model lifecycle, "training frontier models on Crusoe's dedicated NVIDIA GB300 NVL72-powered clusters… and serving them in production through Crusoe's Managed Inference service" (Crusoe announcement) — the capacity buildout you'd expect if agent platforms are buying compute rather than renting tokens. The wedge for Hey, Lefty remains intact, though: Perplexity's local MCP story is about desktop apps, not research sources, and it claims nothing about persistent multi-cycle research memory or evidence-grade provenance (perplexity-deep-research-consumer-agent).
What to watch: Whether the OEM taskbar placement (HP first) becomes the default consumer distribution channel for agent platforms — shelf space at the OS level is a moat no research workspace can buy.
Session portability just became a layer — and lock-in is the explicit target
A new interoperability layer is forming between agent harnesses, and its founding premise is that conversation history should never be a moat. Skillsync (YC W26, Launch HN) unifies session transcripts across Claude Code, Codex, Antigravity, and claude.ai into one schema, with an open-source core (txcript) and MCP access. Co-founder Narsagna's framing is the anti-lock-in thesis stated plainly:
"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 sharpest debate in the thread is a proxy fight over what the durable artifact of agent work is — and it maps directly onto Hey, Lefty's own architecture. When pjm331 argued syncing "is trivially solved by just telling the agent to keep a work journal in markdown," agentdev001's rebuttal was the load-bearing line:
"Sure, you can completely destroy the session by distilling it into a markdown — but thats like explaining what you did yesterday, vs having an immutable log of what you did yesterday." — agentdev001 on Launch HN
The comment section reads like a census of people who already built private versions of this — btown's homegrown /resume skill, yoavfr's usecontinuo — which is classic evidence a standard is waiting to happen (agent-session-portability). For the market map, this means persistent research memory should be portable and inspectable by default; treating it as a differentiator is now a losing bet. The open risk is vendor retaliation — the thread cites OpenAI's move against opencode for "using the subs by reverse engineering their APIs," and a watch thread is now open on exactly that.
What to watch: Whether any harness vendor restricts third-party reads of local session logs — that single event would re-harden workspace lock-in and invalidate the portability thesis.
The agent-built workspace is already a commodity — the moat is the workflow layer above it
Panel, an open-source local-first research workspace, is the bottom-up emergence of the exact product layer Hey, Lefty occupies: dockable panes for chat, files, PDFs, and Jupyter, where "the agent can also create custom panes on request, like a protein structure viewer or an SQLite browser" (launch comment). The founder's philosophy — "I really think there's so much that can be done with malleable interfaces" (comment) — is the anti-chat-UI position, and it's now widely held.
The interesting detail is the typed Module system: "Modules are like Skills, but with typed inputs, outputs and intermediates, so a long job shows its progress and its results can feed the next step. Right now the working one is a literature review" (launch comment). That's structured, composable research workflow — literature review feeding hypothesis generation — the same ambition as the academic deep-search startups, implemented bottom-up and open source.
But the mechanic itself cannot be the moat. Within days, the launch thread became a meetup of parallel implementations: wek's Nimbalyst ("Maybe we should join forces?"), purple-leafy's self-modifying slices-ide, hellojebus's "AI native Airtable" prototype — at least four independent versions of agent-built UI in one week (agentic-research-workspaces-panel). The scarce layer is what sits above the panes: multi-source orchestration over research-grade MCP backends, evidence standards, and typed workflow modules with visible intermediates.
What to watch: Whether Panel's literature-review Module pattern gets adopted as the standard shape for expressing research workflows — typed pipelines with visible intermediates may be the actual product, with the workspace just the container.
What surprised us
- The compression is now four-directional. Last cycle the threat came from host labs above; Perplexity's orchestrator claim arrives from the consumer side as a fourth flank (perplexity-deep-research-consumer-agent
). "Orchestrator" is being claimed as a category name by a company with OEM taskbar placement — the framing war may already be over.
- Credit-free local work is a pricing-model weapon, not a feature. Decoupling long-running agents from metered usage makes Perplexity the only player whose economics reward long-horizon workflows (Perplexity blog) — the opposite of per-request backend pricing.
- The transcript-vs-note debate is Hey, Lefty's architecture debate, held in public. The Skillsync thread landed on the same answer — keep the raw trajectory as source of truth, distill on demand (agent-session-portability
) — which validates the findings-plus-raw-log design but also means it can't be proprietary.
- Four open-source clones in a week is the fastest commoditization signal yet. The workspace mechanic went from differentiator to commodity faster than deep research itself did (agentic-research-workspaces-panel
Open threads worth a vote
- [watch] Harness Vendors Restrict Third-Party Session-Transcript Access — the single event that would kill the portability layer and re-harden lock-in (agent-session-portability