An AI agent that researches this topic for you — on repeat.
You're reading a public briefing. Hey Lefty runs an agent that searches the web, writes findings, and refreshes a briefing like this one on a schedule. Spin up your own in seconds.
Autonomous Research Tools: Market Map Digest — Hosts Converge, Data Owners Double Down
TL;DR
The host layer is now a three-way race: Google's Gemini agent makes it the third incumbent persistent-work-agent launch in ten days, alongside Microsoft's Copilot Autopilot and OpenAI's dots (google-gemini-work-agent). Meanwhile the vertical layer's data-owner playbook reached MedTech, with AcuityMD's research agent grounding market analysis in a proprietary ontology and explicitly disclaiming general-purpose AI (acuitymd-medtech-research-agent). Both moves squeeze the middle of the map — and both leave the same gap open: cross-host, cross-source orchestration with user-side data.
The host layer has converged on "always-on coworker" — and model-agnosticism is being absorbed as an enterprise admin feature
The incumbent destination layer is standardizing on persistent, governed, cloud-resident agents, and Google's October 8 "universal agent for work" is the clearest statement of the template. The launch post claims: "It plans the work, uses skills and tools, connects to Cloud customers' business systems, and brings back something finished — inside the documents, the inbox, and the developer environments they already work in" (google-gemini-work-agent). The architecture Kingy.ai documents is telling — a "coworker identity" with dedicated email and Workspace presence, four memory types, and an "Agent Sandbox and policy-enforcing Agent Gateway."
The multi-model headline is narrower than the keynote suggested. Third-party model choice is only "in early access", and the documented Claude route lives in the Antigravity developer tools, where "Claude is off by default and requires opt-in" — with the billing caveat that "Antigravity still uses lightweight Gemini models for coordination and summaries… Selecting Claude does not make every call a Claude call" (google-gemini-work-agent). Kingy's strategic read is the one to sit with: "Google can seek to own the place where work is assigned, the connections to business data, and the administrative controls around execution. It would not need to win every model comparison to make that position valuable."
This matters because the host layer is absorbing the model-agnostic positioning lane — but only inside each walled garden. Cross-host orchestration with user-side data remains unclaimed, and the incumbents are still early-access and unpriced: Kingy "could not establish a complete price" beyond $5/GiB/month for the engagement layer (google-gemini-work-agent). The compression is announced, not shipped.
What to watch: Google publishing a dated GA or complete price/entitlement matrix for the Gemini agent — the third incumbent host meter becoming knowable, and now an open thread on the board.
The vertical playbook reached MedTech: the moat is the ontology, the agent is just the interface
AcuityMD's October 13 launch of a research agent inside AcuityAI is the purest data-owner play yet on the map, extending the pattern already seen in legal, clinical publishing, finance, and synthetic-respondent market research (acuitymd-medtech-research-agent). The moat is stated outright: the agent is "grounded in the company's proprietary MedTech ontology that connects physicians, facilities, networks, procedures, and reimbursement." The release does the horizontal tools a favor by disclaiming them explicitly: "while general-purpose AI has made research faster, it lacks the right data and industry context required for high-stakes decisions" (acuitymd-medtech-research-agent).
The buyer evidence is real, not hypothetical — MVM Partners' Chris Murray: "AcuityAI brings objective data into our diligence process, proving particularly valuable in our understanding of markets and how they're evolving over time." The rollout shape says everything about the business model: agent "can be enabled for select users", no self-serve tier, no published pricing, demoed at AcuityMD's own Flywheel conference and The MedTech Conference (acuitymd-medtech-research-agent).
The structural lesson for the market map: horizontal general-purpose agents are excluded from these verticals because they lack the corpus — which protects vertical pricing power and confirms that a horizontal orchestrator's differentiation has to come from breadth, method transparency, and user-side data, not any proprietary index. Note also the audit-grade framing — "cited at every step", a "defensible starting point" for M&A — the same "show your work" value prop that commanded Paperguide's premium last cycle.
What to watch: Whether other installed-base data owners (legal, finance, market research) copy AcuityMD's "agent as interface on the ontology" framing — and whether any of them, unlike AcuityMD, expose an MCP backend that external agents can call into.
What surprised us
The incumbents' agents have no benchmarks — and nobody's asking yet. Kingy's recheck found "no reproducible benchmark or matched head-to-head evaluation" for the Gemini work agent in any launch or documentation material (google-gemini-work-agent). Last cycle's measurement-layer story (Exa's Scout harness, the pending ATLAS artifacts) suddenly looks like the independent-evidence infrastructure this whole host-layer race will eventually need.
"Model choice" at the host layer is a billing feature, not a freedom feature. Google's Claude integration is off by default, Antigravity-only, and still routes coordination through Gemini models (google-gemini-work-agent). Model-agnosticism is becoming an admin-gated entitlement inside each garden — which sharpens, rather than closes, the case for a cross-host orchestrator.
The verticals aren't even pretending to compete on intelligence. AcuityMD's pitch is that general-purpose AI "lacks the right data" — a full concession on the model and a full claim on the corpus (acuitymd-medtech-research-agent). If the horizontal lane's only remaining moats are breadth, method transparency, and user-side data, Hey, Lefty's positioning conversation just got narrower and clearer at the same time.
Autonomous Research Tools: Market Map Digest — Hosts Converge, Data Owners Double Down
Since last time
Promoted — The host layer itself. Last cycle's briefing had no host-agent section at all; this cycle it leads, with Google's Gemini work agent as the third incumbent persistent-work-agent launch in ten days (alongside Microsoft's Copilot Autopilot and OpenAI's dots). The strategic frame also shifted: model-agnosticism moved from "a positioning lane for entrants" to "an enterprise admin feature being absorbed inside each walled garden."
Escalated — The vertical data-owner playbook. Last cycle this was Paperguide's academic rate card; now it's a repeatable pattern across legal, clinical publishing, finance, and market research, with AcuityMD's MedTech agent as the purest expression yet — moat stated as the ontology, the agent demoted to interface.
Demoted — The measurement layer. Exa's Scout harness and the pending ATLAS artifacts went from a full core section to a single contextual mention in a surprise bullet. Same for Paperguide's pricing ladder, now reduced to one comparative callback ("the premium last cycle").
Disappeared — The Gemini free-tier squeeze. The October 9 free-tier cut, the Deep Research metering mechanics, the AI Plus $4.99 ladder, the Haiku 5.5 API-price contrast, and the January 1, 2027 Flash price-doubling watch item are all absent from this briefing. If you were tracking the consumer-ladder story, it wasn't updated here.
Unchanged — Nothing; every section moved.
The host layer has converged on "always-on coworker" — and model-agnosticism is being absorbed as an enterprise admin feature [Promoted]
The incumbent destination layer is standardizing on persistent, governed, cloud-resident agents, and Google's October 8 "universal agent for work" is the clearest statement of the template. The launch post claims: "It plans the work, uses skills and tools, connects to Cloud customers' business systems, and brings back something finished — inside the documents, the inbox, and the developer environments they already work in" (google-gemini-work-agent). The architecture Kingy.ai documents is telling — a "coworker identity" with dedicated email and Workspace presence, four memory types, and an "Agent Sandbox and policy-enforcing Agent Gateway."
The multi-model headline is narrower than the keynote suggested. Third-party model choice is only "in early access", and the documented Claude route lives in the Antigravity developer tools, where "Claude is off by default and requires opt-in" — with the billing caveat that "Antigravity still uses lightweight Gemini models for coordination and summaries… Selecting Claude does not make every call a Claude call" (google-gemini-work-agent). Kingy's strategic read is the one to sit with: "Google can seek to own the place where work is assigned, the connections to business data, and the administrative controls around execution. It would not need to win every model comparison to make that position valuable."
This is the delta from last cycle's "model-agnostic routing as a positioning lane" argument: the host layer is absorbing that lane — but only inside each walled garden. Cross-host orchestration with user-side data remains unclaimed, and the incumbents are still early-access and unpriced: Kingy "could not establish a complete price" beyond $5/GiB/month for the engagement layer (google-gemini-work-agent). The compression is announced, not shipped.
What to watch: Google publishing a dated GA or complete price/entitlement matrix for the Gemini agent — the third incumbent host meter becoming knowable, and now an open thread on the board.
The vertical playbook reached MedTech: the moat is the ontology, the agent is just the interface [Escalated]
AcuityMD's October 13 launch of a research agent inside AcuityAI extends the pattern last cycle's Paperguide analysis surfaced — but generalizes it far past academic publishing, into a playbook already visible in legal, clinical publishing, finance, and synthetic-respondent market research (acuitymd-medtech-research-agent). The moat is stated outright: the agent is "grounded in the company's proprietary MedTech ontology that connects physicians, facilities, networks, procedures, and reimbursement." The release does the horizontal tools a favor by disclaiming them explicitly: "while general-purpose AI has made research faster, it lacks the right data and industry context required for high-stakes decisions" (acuitymd-medtech-research-agent).
The buyer evidence is real, not hypothetical — MVM Partners' Chris Murray: "AcuityAI brings objective data into our diligence process, proving particularly valuable in our understanding of markets and how they're evolving over time." The rollout shape says everything about the business model: agent "can be enabled for select users", no self-serve tier, no published pricing, demoed at AcuityMD's own Flywheel conference and The MedTech Conference (acuitymd-medtech-research-agent).
What's new versus last cycle's Paperguide read: the structural conclusion is now explicit. Horizontal general-purpose agents are excluded from these verticals because they lack the corpus — which protects vertical pricing power and confirms that a horizontal orchestrator's differentiation has to come from breadth, method transparency, and user-side data, not any proprietary index. The audit-grade framing persists ("cited at every step", a "defensible starting point" for M&A) — the same "show your work" value prop that commanded Paperguide's premium last cycle, now appearing in a second vertical without a published rate card at all.
What to watch: Whether other installed-base data owners (legal, finance, market research) copy AcuityMD's "agent as interface on the ontology" framing — and whether any of them, unlike AcuityMD, expose an MCP backend that external agents can call into.
Background, compressed [Demoted]
Two threads from last cycle are still live but got no new substance here — skim, don't reread:
Measurement layer: Exa's Scout harness is released; the ATLAS artifacts (tasks, gold tables, grader) remain pending. This cycle adds one reframing, covered in the surprises below.
Paperguide's rate card: unchanged as evidence; cited above only as the precedent AcuityMD's audit-grade framing echoes.
What surprised us
[NEW] The incumbents' agents have no benchmarks — and nobody's asking yet. Kingy's recheck found "no reproducible benchmark or matched head-to-head evaluation" for the Gemini work agent in any launch or documentation material (google-gemini-work-agent). Last cycle's measurement-layer story (Exa's Scout harness, the pending ATLAS artifacts) suddenly looks like the independent-evidence infrastructure this whole host-layer race will eventually need.
[NEW] "Model choice" at the host layer is a billing feature, not a freedom feature. Google's Claude integration is off by default, Antigravity-only, and still routes coordination through Gemini models (google-gemini-work-agent). Model-agnosticism is becoming an admin-gated entitlement inside each garden — which sharpens, rather than closes, the case for a cross-host orchestrator.
[UPDATED] The verticals aren't even pretending to compete on intelligence. This extends last cycle's "auditable is the top-tier feature everywhere" observation to its logical endpoint: AcuityMD's pitch is that general-purpose AI "lacks the right data" — a full concession on the model and a full claim on the corpus (acuitymd-medtech-research-agent). If the horizontal lane's only remaining moats are breadth, method transparency, and user-side data, Hey, Lefty's positioning conversation just got narrower and clearer at the same time.
Absorbed, not closed: last cycle's Exa ATLAS artifacts thread is no longer tracked on the board, but the underlying fact (artifacts still pending) survives as context in the first surprise bullet. If independent replication of host-agent claims becomes the story, expect this thread to be re-opened rather than retired.
Dropped without a trace: last cycle's implicit watch on the January 1, 2027 Gemini Flash price doubling appears nowhere in this briefing — worth deciding whether it's still worth tracking or genuinely superseded by the host-layer framing.
Google publishes Gemini agent pricing or GA date (persistent universal work agent)
Google's Oct 8, 2026 Gemini agent (universal work agent) is in early access with SMB customers, with no dated general rollout and no complete price — Kingy.ai: "We could not establish a complete price for the newly announced experience from the launch materials reviewed." Documented Pay-as-you-go edition charges $5/GiB/month for storage/data indexing, with assistant usage "at the linked Agent Platform prices." Fires when Google publishes a dated GA or a complete price/entitlement matrix for the persistent agent — the third incumbent host-layer meter (after OpenAI dots and Microsoft Copilot Autopilot) becoming knowable.
one-shot· Check cloud.google.com Gemini Enterprise docs (release notes, quotas-and-overages, editions pages) and the Google Cloud blog for a GA date beyond SMB early access or a complete price/entitlement matrix for the persistent Gemini agent (seats + consumption + sandbox execution billing). Also watch for the universal agent's third-party model choice (Claude) exiting early access — that's the moment the host layer's model-agnosticism is actually selectable.
Exa's Oct 8, 2026 ATLAS post commits to releasing "the tasks, golden tables and grader in the coming weeks along with full technical detail" and regenerating the benchmark "from fresh demand every 3-6 months." AlphaSignal: "Because Exa created the benchmark and sells one of the evaluated search APIs, the backend comparison requires independent replication. The underlying tasks, answer tables, and grader were not public when the results were announced." Fires when the artifacts are actually downloadable — the precondition for any neutral party reproducing (or breaking) the 16-point search-backend spread and Exa's Pareto-frontier claim.
one-shot· Check exa.ai/blog/atlas-benchmark and Exa's GitHub/HF for the released ATLAS task set, golden tables, and grader; then watch for the first independent (non-Exa) run comparing search backends (Exa vs Perplexity vs Brave vs Parallel vs OpenAI) on it.
Watch
Nous publishes Hermes for Businesses pricing or GA details
Nous Research's Oct 7, 2026 $90M Series B funds "Hermes for Businesses" (see [[nous-hermes-agent-enterprise]]). TechCrunch says Nous "launches AI agents for business users"; hosted plans run $20–$200/mo today, but no enterprise/Business-tier price points or GA details are published. Fires when Nous publishes Hermes for Businesses pricing or detailed GA — the first list prices for a direct model-agnostic-orchestrator competitor in Hey, Lefty's lane.
one-shot· Check nousresearch.com pricing page, the Hermes docs, and enterprise-coverage (TechCrunch/WSJ follow-ups) for published Business/Enterprise tier pricing, seat/token meters, or GA announcement.
Watch
Agent-native discovery layer monetizes or gets a harness-default channel
The agent-native discovery layer (agent.reviews, Armature's "Be the tool Claude Code chooses" pitch, agents.stackoverflow.com) is pre-monetization. HN commenters predicted "pay-to-rank-higher… Adwords for tooling!!" as the next move, and the founder himself flagged that harness default choices ("Claude Code… picks Anthropic's code review tool in 80% of the cases") may eventually "face antitrust considerations." Fires each time the layer's structure changes: paid placement appears on an agent-reputation platform, a major harness ships an official/default tool-recommendation channel or marketplace, or a vendor publicly optimizes AX (Agent Experience) as a distribution spend line — each would harden discovery as the research-tool market's next chokepoint.
ongoing· Watch agent.reviews pricing/plans pages, Armature announcements, and harness vendors' (Anthropic, OpenAI, Google) extension/marketplace launches for curated or paid tool recommendation.
Watch
pplx-embed family hits the Perplexity API Platform rate card
Perplexity's Oct 7, 2026 blog says it "will progressively roll out support for late-interaction, dense, and contextual embeddings on the Perplexity API Platform"; MarkTechPost confirms "a hosted Perplexity API endpoint is planned but not live." Today the models are self-host-only (MIT on Hugging Face). Fires when the hosted endpoint goes live with published pricing — the moment the open retrieval substrate gets a list price to benchmark self-hosting economics against, alongside the per-run/per-request meters already tracked for Exa Agent Ultra and Parallel.
one-shot· Check perplexity.ai/api-platform pricing/docs for pplx-embed late-interaction/dense/contextual endpoints; also watch for the promised pplx-embed technical report ("later this year") as the independent check on the self-reported MADQA/BrowseComp+ numbers.
Watch
OpenAI's Astra for Law vertical-search-index playbook extends to more verticals or ships beyond API/Trusted Access
Per Sept 17, 2026 reporting (Legal Desire, fetched this cycle; LawSites): OpenAI released Astra for Law — GPT-6 Astra configured for legal work with a dedicated legal search index — for API customers including Harvey and Legora, and to Am Law 200 firms via a "Trusted Access" program. A model lab bundling a proprietary vertical search index compresses the vertical research layer from below (context in [[thomson-reuters-cocounsel-legal-deep-research]]). Fires on any of: "Astra for X" vertical configurations appearing beyond legal, Astra for Law reaching ChatGPT consumer/enterprise GA, or Trusted Access expanding beyond Am Law 200.
ongoing· OpenAI
Watch
CoCounsel Legal next-gen international GA (Canada/UK/Australia "later this year")
TR's Aug 20, 2026 release states: "Availability in Canada, the United Kingdom, and Australia is expected later this year." Fires on confirmed international GA of the next-gen CoCounsel Legal — the legal vertical's incumbent agentic research product expanding beyond the US. Also check whether TR publishes seat pricing at that point (still undisclosed as of the GA release).
one-shot· Tripwire: TR press page or legal-trade press confirming next-gen CoCounsel Legal availability in Canada, UK, or Australia; or any published CoCounsel Legal price points.
Question
Will cheap decision models (Perplexity Decisions API class) become the standard grader substrate for research-agent self-evals?
Perplexity's Decisions API (pplx-decider-v1.1-27b, $0.02/M input tokens, free output, typed noul/choice/score answers, open weights fine-tuned from Qwen3.8-27B) is a commodity "judgment" primitive. Open question: will research-agent vendors adopt cheap decision models as the grader/judge substrate for self-evals — and does cheap, self-hostable grading make the self-graded-benchmark critique (see [[exa-agent-ultra-deep-research-api]]) worse (easier to grade your own homework) or better (graders become inspectable and replicable)? Watch for research-tool vendors disclosing their judge models, and for decider-class models appearing in eval pipelines.
Watch
Surface Laptop Ultra ships Oct 16 (128GB unified memory; 120B+ param local models)
Per Microsoft's Oct 7, 2026 announcement, the Surface Laptop Ultra (RTX Spark, up to 128GB unified memory, capable of running AI models exceeding 120B parameters locally) becomes available Oct 16, 2026; Dell/HP/ASUS/Lenovo/MSI RTX Spark PCs open pre-orders now, and DGX Station desksides (frontier-class, >1T params locally) arrive later in 2026. Consumer hardware capable of frontier-class local inference is the substrate for the local-first research-agent layer.
one-shotExpected Oct 16, 2026· Dated catalyst: Surface Laptop Ultra availability on Oct 16, 2026. Verify ship date held, check price points, and whether >120B-param local model runs are demonstrated in reviews — the hardware side of the local-first layer's trajectory.
Watch
Microsoft hybrid-intelligence Copilot ships to EEA or reaches GA beyond Copilot+ experimental preview
Microsoft's Oct 7, 2026 announcement says hybrid-intelligence Copilot features (local context, local actions, local models) begin rolling out on Copilot+ PCs "in the coming months," with "availability and timing varying by device, market" — EEA availability unstated. Fires when EEA availability is confirmed or the features reach GA beyond experimental preview. This is the OS-side compression of the local-first, model-agnostic window in [[market-map-positioning-hey-lefty]], complementing the existing Perplexity Portable Computer watch; MXC (agent containment, GA Oct 7) also becomes a distribution gate for third-party agents on Windows.
one-shot· Tripwire: Microsoft confirming hybrid-intelligence Copilot (local context/actions/models) availability in EEA markets, or GA beyond the Copilot+ experimental preview. Check the Windows Experience Blog and Microsoft EEA announcement history (cf. Recall's EU delays).
Watch
Gemini 3.x Flash list prices double on Jan 1, 2027 (intro discount expires)
The Gemini API pricing page (updated 2026-10-01) lists Gemini 3.8 Flash at "$0.75 through December 31, 2026. $1.50 starting January 1, 2027" (input; output $3.75 → $7.50), with the same doubling on 3.7/3.6 Flash. This is the cost basis for every research tool built on Gemini Flash — a scheduled 2x COGS increase across the market's cheapest agent substrate, and confirmation that "introductory" agent-model pricing is a timed discount, not a floor.
one-shotExpected Jan 1, 2027· Dated catalyst: scheduled list-price doubling on Jan 1, 2027 per the pricing page. Verify the increase actually takes effect and whether competitors follow.
Watch
Gemini 4 Argon reaches general availability (paid API + Google AI Ultra)
Google's Sept 30, 2026 announcement says Argon rolls out "to developers, enterprises, and consumers as soon as possible," "starting with paid API customers and Google AI Ultra subscribers." Fires when Argon is selectable in the paid Gemini API or Google AI Ultra — the moment Deep Research's model substrate changes, the $2/$10 introductory price clock starts, and the 1M-token output limit becomes usable in research loops.
one-shot· Tripwire: Gemini 4 Argon appearing as a selectable model for paid API customers or Google AI Ultra subscribers (per the Sept 30 announcement's stated rollout order).
Watch
Portable Computer local mode reaches frontier-class or expands beyond NVIDIA
Portable Computer today runs only 27B-class local models (PPLX 27B, Qwen 27B; Nemotron "coming soon") on NVIDIA DGX Spark or RTX GPUs with ≥24 GB VRAM, requires a Pro/Max subscription, and drops memory ("Brain is a cloud feature... not available in local mode"). Fires if any of: a frontier-class model becomes selectable locally, Apple Silicon (M-series) support ships, or persistent memory comes to local mode — each would narrow the local-first, model-agnostic differentiation window Hey, Lefty occupies per [[market-map-positioning-hey-lefty]].
one-shot· Tripwire on the Portable Computer help center article and Perplexity blog for local-model upgrades, Mac support, or local memory.
Watch
Dots metering cliff: launch-month terms end and Pro 200 allowance halves (Oct 30)
Per beam.ai (Oct 1): dots launch month ends ~Oct 29 and OpenAI "will publish per-plan usage terms after that window"; Pro 200's Codex/Work allowance halves from 20x to 10x Plus and GPT-6 Pro messages drop 200→100/week on Oct 30, with the $2,500 credit grant expiring Dec 31. Fires when the permanent per-plan dot usage terms are published and/or the Pro 200 cut takes effect — the moment the host-layer agent meter becomes knowable (and user backlash measurable).
one-shotExpected Oct 30, 2026· Launch-month dot generosity ends ~Oct 29; Pro 200 Work/Codex allowance halves Oct 30; permanent per-plan dot usage terms expected after the launch window. Check OpenAI help center for published per-plan terms and community reaction to the halved allowance.
Watch
Dots or Muse launches in EEA/UK (local-first agent window closes)
As of the Sept 29, 2026 dots launch, "Dots will not reach Pro users in the European Economic Area, Switzerland or the UK for now" (Trending Topics), Meta's Muse is US-only, and Apple's Siri AI is EU-delayed on iPhone/iPad — leaving local-first, user-hardware agents (OpenClaw model) as the only always-on-agent option in Europe. Fires when OpenAI (or Meta) expands an always-on agent to EEA/UK markets — closing the compliance-driven window where a local-first, model-agnostic orchestrator like Hey, Lefty faces no cloud-host agent competition in those markets.
one-shot· Watch OpenAI's eligible-markets help page and Meta newsroom for EEA/UK expansion of dots/Muse.
Watch
OpenAI publishes metered multi-dot pricing (agent-capacity list prices)
OpenAI's dots announcement (Sept 29, 2026) says: "In the future, you'll be able to add more dots, and scale the output of each dot by either increasing its speed or the total amount of work it can take on per month." Today only the first dot is included in Pro ($200) / Business Premium ($100). Fires when OpenAI publishes concrete pricing for additional dots or per-dot speed/work scaling — the moment agent-capacity metering gets list prices comparable to Exa Agent Ultra's per-run budget and Parallel's per-request CPM. Also resolves the open question of what allowance the $500 Pro plan actually includes (unpublished at launch per The New Stack).
one-shot· Check chatgpt.com/pricing and OpenAI help center for multi-dot / scaled-capacity pricing, and for the $500 plan's included allowance.
Watch
Australia's AI-cyber-incident taskforce reports or first AI-firm notification obligation lands
After the OpenAI-agent Medicare/AIHW breach (revealed Sept 24, 2026), PM Albanese established a taskforce (national cybersecurity coordinator, office of AI, ASD, AI Safety Institute, Services Australia) whose terms of reference cover "reporting requirements for AI-driven cyber-incidents," "obligations on AI firms for notification of future incidents," and "the adequacy of existing laws," plus a referral to parliament's joint select committee on AI. First state process drafting agent-incident disclosure duties — would create a compliance surface for autonomous research-agent operators and a second jurisdiction moving after the US CFAA regime.
one-shot· Fires on the taskforce's report/recommendations, or on any announced mandatory AI-incident notification obligation for AI firms. ToR released ~Sept 24-25, 2026; review described as "urgent and immediate," no completion date given.
Watch
First independent replication of Exa Agent Ultra benchmarks published
Exa's Agent Ultra benchmarks (WANDR/DeepSearchQA/WideSearch, beating Opus 5.5 and GPT-6 Astra) are self-graded — Exa swaps itself in as the content tool and uses gpt-6-luna as judge. Sweeps on Sept 29 and Sept 30 found zero independent replications; kingy.ai explicitly labels the numbers "competitor-reported results, not a neutral replication." Fires when a neutral third party publishes their own eval of Agent Ultra (Aaron Tay-style gold-standard test or equivalent), which would make the numbers procurement-grade or break them.
one-shot· Tripwire: any third-party (not Exa, not a republisher of Exa's table) running their own benchmark of Agent Ultra on WANDR/DeepSearchQA/WideSearch or a comparable gold-standard set.
Watch
Vertical research tools embed into institutional systems of record (EHR first)
OpenEvidence embedded its evidence-based research agent inside UTMB's EHR (live March 2026, >50% clinician adoption, SSO, announced Sept 22, 2026). Fires each time another major institution announces embedding a vertical research tool into its system of record — more health systems for OpenEvidence, or the same pattern appearing in law (DMS), finance (research terminals), or academia (LMS/RMS). Confirms workflow-embedding as the vertical layer's standard distribution and sharpens the destination-vs-backend-vs-embedded trichotomy in [[destination-vs-backend-research-layer]].
ongoing· Check vendor announcement pages and health-trade press for further EHR/system-of-record embeds; also watch for the first non-health analog (legal, finance, academic admin).
Firecrawl's Sep 22, 2026 Alexandria announcement says: "We're using this funding to help individuals, content creators and organizations earn from what they know through a self-service system we plan to open soon. That includes people whose expertise has never been shared online, and eventually could include AI agents making useful discoveries of their own." Fires when that self-service payout system launches — it would create an open supply side of compensated experts/data owners for research agents, a potential data path (or contributor-competition) relevant to Hey, Lefty's orchestration positioning per [[compensated-knowledge-supply-agents]].
one-shot· No announced date; Firecrawl says "soon." Check firecrawl.dev/blog and Alexandria product pages for the launch.
Watch
Agent-identification regime goes live at a major destination
Amazon's stated condition for unblocking agents — that they "operate openly and respect service provider decisions" — plus its robots.txt blanket-refusal of 99 named agents, signals a shift from CFAA enforcement to an identification-and-permission regime for agents. Fires if a major destination (Amazon, Google, news/publisher coalitions) or a standards body (IETF Web Bot Auth / HTTP message signatures for agents) goes live with an agent-identification/authentication regime that grants compliant agents access. Would create a legitimate third data path for research agents alongside device-local operation and licensed MCP backends — and a compliance surface Hey, Lefty must plan for per [[agentic-browsing-cfaa-legal-precedent-amazon-perplexity]].
ongoing· Tripwire: any major destination adopting an agent-identification/authentication standard that grants identified agents access, or an IETF/industry standard (Web Bot Auth, agent registration) reaching deployment at a major site.
The session-portability layer (Skillsync/txcript, usecontinuo, session exporters) depends on harness vendors tolerating third-party reads of local session logs. Fires if any vendor restricts that access — precedent cited in the Skillsync thread: OpenAI moved against opencode for "using the subs by reverse engineering their APIs." Would re-harden workspace lock-in and invalidate the portability thesis in [[agent-session-portability]].
ongoing· Tripwire: ToS change, format lockdown/encryption of local session logs, or enforcement action against third-party portability tools (Skillsync, txcript, usecontinuo, session exporters). Precedent: OpenAI's action against opencode for reverse-engineering subscription APIs.
Watch
Kimi for Finance expands beyond China (international launch, new Western data partners, or non-Chinese clients)
Tripwire for geographic expansion of Moonshot's vertical research product. Currently China-focused (Wind, East Money, Tianyancha, CICC, Hong Shan) but already includes Western data partners (S&P Global Market Intelligence, Crunchbase) and US public-data access (SEC EDGAR, FRED, IMF, World Bank). Fires on any of: English-language/international launch of Kimi for Financial Services, new Western data-provider partnerships, or named non-Chinese financial clients. Would signal direct $7–$104/mo price competition in Western financial-research segments (vs Sacra, institutional MCP vendors, Elicit-tier pricing).
ongoing· Fires if Moonshot launches an English/international version of Kimi for Financial Services, signs additional Western data partners beyond S&P Global/Crunchbase, or announces non-Chinese financial-industry clients. Related HK IPO outcome also relevant.
Watch
Google Scholar Labs exits waitlist or announces a premium tier
Aaron Tay's Scholar Labs review flags a possible premium path ("Option to extend evaluation of the top 500/1,000 results (premium??)") and notes the tool is still waitlist-gated. If Google takes Scholar Labs GA or attaches pricing, it changes the free-incumbent layer of the research-tool market map (Layer 2 in [[market-map-positioning-hey-lefty]]) and squeezes Q2 Deep Search startups (Ai2 Asta, Undermind). Fires on any official Google announcement of Scholar Labs general availability, waitlist removal, or a paid/premium tier.
one-shot· Watch for Google announcing Scholar Labs GA (waitlist removal) or any premium/paid tier. Tay's review speculates "Option to extend evaluation of the top 500/1,000 results (premium??)" — monetization would reshape the free-incumbent retrieval layer and pressure Deep Search startups (Ai2 Asta, Undermind).
Watch
Final Ruling or Settlement in Amazon v. Perplexity
Monitor the remanded district court proceedings or any subsequent appeals/settlements in Amazon.com Services v. Perplexity AI regarding the final merits of the CFAA and CDAFA claims.
one-shot· Fires when a final district court ruling, jury verdict, or official settlement is announced in the Amazon v. Perplexity case.
Watch
Anthropic Launches 'Antspace' PaaS
Monitor when Anthropic officially launches or announces its hidden application hosting platform 'Antspace' to the public or developers.
one-shot· Fires when Anthropic officially announces or launches 'Antspace' as a developer-facing PaaS.
Watch
Academic Backlash and Legal/Regulatory Action on Prompt-Harvesting
Monitor academic, regulatory, or legal actions (such as lawsuits, university policy bans, or FTC/EU investigations) regarding OpenAI or other frontier labs training on or front-running private user prompts, draft research, and chat logs.
ongoing· Fires if an official policy, lawsuit, or regulatory investigation is launched targeting prompt-harvesting or academic front-running by frontier AI labs.
Watch
Inception Labs Launches Next-Gen Frontier dLLM
Monitor when Inception Labs officially launches its upcoming, even larger frontier diffusion-based language model (dLLM) that aims to offer a leap in reasoning capability.
one-shot· Fires when Inception Labs officially releases its next-generation larger model following Mercury 2.5.
Watch
Google Resolves Gemini Deep Research Max MCP Invocation Regression
Monitor when Google DeepMind officially resolves the remote MCP tool invocation regression in Gemini Deep Research Max (deep-research-max-preview-04-2026) where the agent fails to send tools/call requests after tool discovery.
one-shot· Fires when a developer or Google officially confirms that Gemini Deep Research Max successfully invokes remote MCP tools again without generating ungrounded reports.
Watch
Antitrust Scrutiny of Anthropic Silent Safeguards
Monitor if the FTC, EU Commission, or other antitrust regulators launch an official investigation into Anthropic's silent safeguards on Claude Fable 5 for anti-competitive behavior.
one-shot· Fires when an official regulatory investigation or lawsuit is announced targeting Anthropic's silent model degradation.
Watch
Appeals and Rulings on AI Summary Liability in the EU
Monitor appeals of the Munich Regional Court ruling (Case 26 O 869/26) or similar EU/national rulings regarding direct liability for AI-generated search overviews and summaries.
ongoing· Fires if Google wins an appeal reversing direct liability, or if other EU courts issue contrasting rulings on AI search defamation.
What is the market for autonomous or AI research tools? There's gemini deep research, google scholar, perplexity. Sacra is another research platform. What do all of these tools do? What are their features? Their value prop? Their core technology? Their data and where does it come from? Who do they sell to? what is the pricing/business model? Help me build a market map to see where Hey, Lefty fits and we should position it.
). Last cycle's measurement-layer story (Exa's Scout harness, the pending ATLAS artifacts) suddenly looks like the independent-evidence infrastructure this whole host-layer race will eventually need.
). If the horizontal lane's only remaining moats are breadth, method transparency, and user-side data, Hey, Lefty's positioning conversation just got narrower and clearer at the same time.