Autonomous Research Tools: Market Map Digest — The OS Moves, the Host Builds an Economy, and Provenance Becomes the Only Trust Left
TL;DR
This cycle's headline is compression from below and commoditization from above: Microsoft shipped local context, OS-wide agent actions, and an agent-containment gate (MXC) that turns Windows itself into a competitor to the local-first layer — and a distribution tollbooth for everyone on it (microsoft-copilot-local-agent-stack). Meanwhile Perplexity is assembling the full host-layer economy — metered compute, Skills, a bank channel, team-context connectors, and now a metered "decision primitive" with open weights (perplexity-deep-research-consumer-agent
). And the trust substrate underneath all of it — style, watermarks, token statistics — is collapsing in both directions, which makes process evidence (citations, replayable runs) the last durable differentiator (ai-expertise-proxies-goodhart
).
Microsoft turns the operating system into a research-agent platform — and a gate
The local-first, model-agnostic window is being compressed from the OS side, not just the host-app side. On October 7, Microsoft announced Copilot gains three local capabilities on Copilot+ PCs — local file context with user permission, OS-wide actions ("organizing files, assessing device diagnostics, troubleshooting issues, coding, and carrying out workflows directly on the device"), and on-device models — with MAI Code 1.1 Flash running at 3-bit on device and a 256K context window (microsoft-copilot-local-agent-stack). The same day, Microsoft Execution Containers (MXC) went GA as a policy-driven containment layer, and the framing is the tell:
"the containment layer keeps policy outside the agent workload's control, so the agent or its generated code cannot grant itself additional access." — Microsoft Gives Copilot Local File Context and OS-Wide Actions – Unite.AI
MXC compliance is already supported by OpenAI's Codex, GitHub Copilot, Replit, and others, with Perplexity and Meta's Muse queued as planned supporters — meaning it is becoming a distribution requirement for any agent that wants to act on Windows (microsoft-copilot-local-agent-stack). For the market map, two consequences: "runs locally on your machine" stops differentiating against the default OS itself, and a new compliance surface — MXC — joins the legal-regime watches already on the board. The durable differentiators for a player like Hey, Lefty narrow to model-agnosticism (Microsoft routes to its own MAI/Nemotron/DeepSeek stack via HydraFusion) and cross-harness portability.
What to watch: EEA availability of the hybrid-intelligence Copilot features — unstated in the announcement ("availability and timing varying by device, market") — and the Surface Laptop Ultra's October 16 ship date, which puts 120B+ parameter local inference in consumer hands (microsoft-copilot-local-agent-stack).
Perplexity is building the host layer's full economy — and claiming your slot in marketing copy
Perplexity's October wave (Oct 1–7) reads like a checklist of every layer a research orchestrator would otherwise have to build. The Decisions API sells judgment as a metered primitive — pplx-decider-v1.1-27b at $0.02/M input tokens, free output, typed probability answers, with open weights on Hugging Face (perplexity-deep-research-consumer-agent). Distribution arrived via American Express: ten business Skills for U.S. Amex Business Card members, gated on an active Enterprise subscription and a read-only Plaid card link. And the context layer arrived via Rings AI's relationship-graph connector, which Perplexity's own boilerplate frames in terms that should stop a positioning exercise cold:
"a model-agnostic orchestrator of AI agents across tools, files, chips, devices, and more" — Rings AI and Perplexity Partner to Bring Team Relationship Intelligence into Perplexity Computer — PR Newswire
That is the exact slot the market map assigns to Hey, Lefty, now claimed in hyperscale marketing language (perplexity-deep-research-consumer-agent). But the decision primitive cuts both ways: it's a building block any research agent could adopt as a grader/evaluator, and it commoditizes one more layer. The open question — flagged as a thread this cycle — is whether cheap, self-hostable decision models become the standard judge substrate for research-agent self-evals, and whether that makes the self-graded-benchmark critique worse (easier to grade your own homework) or better (graders become inspectable).
What to watch: Whether research-tool vendors start disclosing their judge models — the first movers on judge transparency will either defuse or weaponize the self-graded-benchmark critique.
The proxies for expertise are breaking in both directions — provenance is what's left
The trust layer under the entire research-tool market is eroding simultaneously on the detection side and the dismissal side. On October 7, WildSEO launched TextGrain Remover, a free browser tool that strips "synthetic watermarks, invisible Unicode artifacts, and statistical token signatures from AI-assisted copy," including a rewrite mode that breaks "predictable n-gram token frequencies while strictly retaining original names, numbers, citations, and factual claims" (ai-expertise-proxies-goodhart). Meanwhile, on the HN thread where human work gets dismissed as machine-written:
"Whenever I see those threads, they usually don't discuss the merits of the writing. They claim that AI must have been used, so it's not even worth reading." — Comment by aprilthird2021 / reply by pitkali
When style-based proxies false-positive on humans and technical proxies can be stripped for free, detection becomes an arms race with zero-cost tooling — and process evidence (citations, replayable runs, source chains) becomes the only trust signal that can't be laundered (ai-expertise-proxies-goodhart). This reinforces the audit-grade positioning already validated by NVivo's $770/year pricing last cycle: the verification axis isn't a feature, it's the surviving moat.
What to watch: Whether answer engines and publishers start weighting watermarked versus scrubbed copy differently — the first credible signal of how search and discovery actually treat provenance.
What surprised us
- The containment gate shipped before the agents did. MXC went GA the same day as the Copilot local features, with seven third-party agents already supporting it and four more queued (microsoft-copilot-local-agent-stack
) — Microsoft is building the tollbooth before the traffic, which is the opposite of how platform gates usually arrive.
- Perplexity is quoting your positioning back at you. "Model-agnostic orchestrator" is now Perplexity's own boilerplate (perplexity-deep-research-consumer-agent
- The laundering tool preserves citations. TextGrain's rewrite mode explicitly retains "names, numbers, citations, and factual claims" (ai-expertise-proxies-goodhart
) — even the strippers know that the citation chain is the part worth keeping, which is itself evidence for where the trust layer lives.
Open threads worth a vote
- Will cheap decision models become the standard grader substrate for research-agent self-evals? — steer whether we track judge-model disclosure as a standing signal across vendors.
- [watch] Microsoft hybrid-intelligence Copilot ships to EEA or reaches GA beyond Copilot+ experimental preview — the OS-side compression of the local-first window; steer follow-up priority on the topic page.