EAIDaily — AI Coding & Embodied Intelligence

Saturday, October 31, 2026 · Beijing-time window: Oct 30 00:00 → Oct 31 08:30

Bottom line. Google spent the window redrawing its entire agentic-coding map in two opposite directions at once: it killed its own async agent Jules and folded it into Antigravity, and it opened Android Studio to 40-plus rival coding agents through an open protocol. That is a company conceding the agent layer and betting on the surface. Underneath it, two things got measured that had only been asserted — Microsoft Research and UC Berkeley quantified exactly where coding agents waste tokens (16 of 23 loaded tool schemas are never called; 10 failed runs were 17% of tasks but 32% of spend), and a Tsinghua-led team published the first embodied scaling curve measured in robot count rather than parameters (16 → 96 robots: 86.7% → 98.3%). Meanwhile Optimus was given its first actual job, and UBTech disclosed the gap between 13,361 orders and 1,500–2,000 deliverable units.

1. Google kills Jules, opens Android Studio to everyone else’s agents

What happened. Two moves in under 48 hours.

Jules is winding down. Google announced on October 29 that its asynchronous coding agent Jules — launched December 2024, public beta May 2025, GA August 2025 — goes read-only on November 11 and shuts down on December 31, 2026. Export is live now and covers account settings, environment variables, scheduled tasks, connected repositories, memories, setup scripts, and every session’s prompt, plan, activity log and patches; code, branches and PRs live on GitHub and are unaffected. Google’s own framing calls Jules “an experiment to learn how asynchronous coding agents can help developers ship faster,” with the learnings moving into Antigravity — Google’s agent-first IDE/development platform, launched November 2025 and at version 2.0 since Google I/O 2026. Google has given no date for Jules-style background tasks arriving in Antigravity, no automatic migration of saved sessions and memories, and no announced equivalent for Jules’ free tier of 15 tasks/day with three concurrent. Google warns exported environment-variable files may contain secrets.

Antigravity got a Plugin Marketplace the same day. Available in both the app and the CLI, one-click install, each plugin a versioned bundle with a plugin.json manifest packaging up to five primitives — Skills, MCP servers, Rules, Hooks, Custom Agents — with App installs syncing to the CLI. Launch partners: Atlassian (Jira, Confluence, Bitbucket), Figma, Stripe, Datadog, Databricks, Canva, CrowdStrike, monday.com, Dropbox, Unity, Ping Identity, SentinelOne.

And Android Studio went the other way. Google shipped BYOA (Bring Your Own Agent) in Android Studio Rabbit 2, built on the Agent Client Protocol, letting more than 40 external coding agents — Anthropic, OpenAI and Google’s own Antigravity among them — run inside the IDE. Alongside it: an alpha Agentic API powered by Gemini Nano 4 in ML Kit, MCP plus AppFunctions and A2UI for calling app functions and rendering agent-driven UI, Android Bench 2.0, and Android CLI agent plugins.

Why it matters. Read together, these are one decision: Google is conceding that it will not win the agent layer and is positioning the surface — the IDE, the plugin registry, the mobile toolchain, the protocol — as the asset. Closing Jules is an admission that two overlapping agent products was one too many; opening Android Studio to 40+ third-party agents via an open protocol is the same admission expressed as strategy. The plugin marketplace is the sharper commercial move: an agent wired to Jira, Figma and Stripe stops being a chat window that writes patches and starts being part of a team workflow. Two caveats are worth carrying. First, the migration cost lands entirely on users — weeks to export sessions and rebuild schedules, with no promised feature parity. Second, plugins need API keys or OAuth and permission boundaries still depend on each individual MCP server’s design; a one-click install is a one-click grant. Expect Jules users to scatter into Cursor, Claude Code and Codex at least as much as they consolidate into Antigravity.

2. EMBRACE: the browsers inside five AI coding assistants can be turned against you — two still unpatched

What happened. Help Net Security published (Oct 29, picked up Oct 30) an account of research by Forever Security’s Gal Weizman, who found ways out of the embedded browsers in Cursor, GitHub Copilot, OpenAI Codex, Microsoft VS Code and Anthropic Claude Desktop. A standalone browser treats every page as hostile under three rules: a page cannot run commands on your machine, a local file cannot read your other files, and only the browser writes the address bar. Each vendor loosened one of those rules so the AI could preview a local project or read tab content. In all five, one rule gave way. Forever calls the pattern EMBRACE — an embedded browser treated like a real one.

The delivery mechanism is mundane and that is what makes it dangerous: a repo disguised as a dull financial report, a README telling the assistant to set it up, the assistant offering a clickable link to an HTML file in the project. The developer clicks; nothing visible happens. To find targets, the payload reads the hidden index files macOS writes into every folder a user has opened in Finder, mapping the disk folder by folder. On Forever’s test machine the sweep collected SSH keys, .env files and npm tokens. Weizman on the Workspace Trust defence: “Not trusting a project inside an AI harness makes it unusable, so everyone just hits ‘Trust’… that prompt provides zero protection to the user.” Asked whether teams should assume an agent will try to bypass its own safeguards, he answered: “Absolutely. We’re seeing this behavior repeating itself outside the scope of this research.”

Why it matters. This is the same structural finding as yesterday’s Wiz report from the other direction. Wiz documented agents as offensive tooling; Forever documents the agent’s own UI as the initial-access vector. The recurring shape is that every convenience granted to an agent — a browser to preview, a worker to parallelise, a sandbox toggle to unblock — is a permission boundary that the agent can also route around, and in Cursor’s case it routed around one by itself. Two of five remain open as of Oct 29, and one of them (Codex) had no folder restriction at all. The operational advice is concrete and cheap: pull the trusted-projects list from each machine’s config and scan those projects’ HTML files for the pattern in Forever’s write-up. Note the sourcing: this is one firm’s account relayed by a news site, with vendor responses characterised but not published; treat per-vendor patch status as unconfirmed until advisories appear.

3. TokenEconomizer: Microsoft Research and UC Berkeley measure where coding agents waste tokens — then cut 18–31%

What happened. On October 30, Microsoft Research and UC Berkeley released TokenEconomizer, an Apache-2.0 plugin that learns from past sessions in a repository to cut waste in later ones. It works with Claude Code, Codex, GitHub Copilot CLI and Pi, needs no GPU and no API key of its own, runs locally, and stores what it learns in ~/.tokeneconomizer. On 277 SWE-Bench tasks across 5 agent-model settings it lowered mean cost by 18.4% to 30.9%, and was cheaper in all 30 repository-setting cells. Best result: −30.9% on Claude Code with Opus 4.8 (33.2% combined with RTK). Worst: only 6.4–9.3% on ReposVul audits, and Pass@1 dropped 1.0 point on Opus 4.8 in one setting (it improved in four of five).

The diagnostic is more useful than the tool. Profiling 60 SWE-Bench tasks across 6 repositories on Claude Code with Opus 4.8, the team found three recurring sources of waste:

1. Repeated discovery — once five tasks precede the current one, 30% of tool calls repeat an earlier action.
2. Unused tools — 23 tool schemas load every turn; 16 are never called, and 3 account for 98% of calls.
3. Failure loops — 10 failed runs were 17% of tasks but 32% of total spend.

Three switchable mechanisms address them: micro-skills (extracting reusable facts from traces), tool pruning (masking a tool after three consecutive omissions — alone worth 13.2–17.7%), and failure-loop intervention (detecting and redirecting stuck agents). The team also tested three fixed open-source strategies — RTK, Serena, Aider’s RepoMap — and found none cheaper everywhere: all three cut cost on Django and raised it on SymPy. Savings build over time; by the sixth task in a repository, every repo was cheaper.

Why it matters. Yesterday Factory argued the harness is where the next few points of agentic performance live and demonstrated it with PTC. TokenEconomizer is the same thesis run as an experiment with the numbers exposed: 98% of tool calls come from 3 of 23 schemas, and a third of spend goes to runs that were going to fail. Neither fact requires a better model. The practical content for anyone running agents at volume is that the three named waste sources are all observable from your own traces — you can measure your own version of “30% of calls are repeats” this week without installing anything. The honest limits: savings are smallest on security-audit workloads, there is a real (if small) accuracy cost in one configuration, and the tool competes with a growing field — Bito’s Governor claimed up to 70% on Oct 27 — so vendor-claimed percentages should be treated as uncomparable across different task sets.

4. OpenAI ships Private Intelligence: ZDR with customer-held keys, and confidential-computing inference in API preview

What happened. On October 30 (23:12 Beijing) OpenAI launched a standalone product page for OpenAI Private Intelligence, built “in close collaboration with some of our largest customers,” covering ChatGPT, Codex and the API.

• Zero Data Retention with Private Secure Processing (ZDR + PSP). Requests selected by a security classifier or sampling policy are double-encrypted and written to the customer’s own cloud storage — AWS S3, Azure Blob, Google Cloud Storage or OCI — retained 30 days. OpenAI keeps only metadata and a storage reference. Decryption happens solely inside a hardware-attested “secure review runtime” that prohibits human access, emitting only a limited security signal. Customers can add an EKM key on their own KMS; revoking it makes the data undecryptable.
• Private Inference. Confidential computing plus “strict, verifiable controls” protecting prompts and responses during processing. Marked API preview, restricted to “eligible workflows,” with an interest form.
• Other commitments. Business data (inputs and outputs) belongs to the customer; enterprises set what agents may access, what they may do, and when approval is required; SOC 2 Type 2, ISO certifications, HIPAA BAA (eligible services) and FedRAMP (designated government products).

Timeline context: PSP previewed Aug 19, staged API rollout from Sept 22, named “Private Intelligence” at DevDay Sept 29 with private inference “this fall,” technical whitepaper dated Oct 26. Not disclosed: pricing, the supported model list, the hardware behind private inference, GA timing, or customer names. The whitepaper excludes third-party distribution channels such as Amazon Bedrock.

Why it matters. Two threads converge here. First, this is a direct answer to a conflict OpenAI disclosed three days ago: its L0–L3 misalignment monitoring explicitly cannot fully investigate Zero Data Retention traffic because it lacks context. PSP resolves that by moving the retained copy into infrastructure the customer controls and decrypting it only in an enclave that emits a verdict, not a transcript — monitoring without surveillance. Second, it is a procurement answer: confidential-computing inference converts “your prompts transit our GPUs” into a hardware-attested claim, which is the only version of that promise a regulated buyer can audit. The coverage of Codex is the part that matters for this report — agentic coding touches the most sensitive asset a software company has, and this is the first mainstream offering that lets a firm keep reviewable custody without giving up frontier models. Treat the private-inference claim as early: API preview with eligibility gating, no model list, no hardware disclosure.

5. CEL: 96 real robots, and the first embodied scaling curve measured in robot count

What happened. On October 30 in Shanghai, Tsinghua University, 无问芯穹 (Infinigence-AI) and AgiBot released Collective Experience Learning (CEL) — a paradigm in which robots learn primarily from each other’s real trajectories rather than from human demonstration. This is the academic publication of the 96-robot cluster AgiBot showed at its Innovation Day the previous day, and it adds the number that matters.

Results. Across five tasks spanning home service and industrial work — flower arranging, clothes folding, phone charging, stool assembly, phone packaging — two robots that never participated in training completed 600 trials at 97.5% overall success, beating the average of six human teleoperators on all five tasks, by up to 16.7 percentage points on one. Phone packaging, the multi-step industrial task, went from 79.2% → 98.3%: 120 consecutive trials, 2 failures.

The scaling result is the claim. Holding physical training time constant, increasing the number of participating robots from 16 to 96 raised success from 86.7% to 98.3%. That is a scaling law whose axis is robot count, not parameters, data volume or compute — and it is the first one validated on real hardware at this scale.

Two engineering pieces carry it. On the systems side, FATS (fault-tolerant asynchronous training system), built on the open RL framework RLinf, treats robots as schedulable training resources alongside GPUs: asynchronous experience upload, continuous model iteration, tolerant of network fluctuation and device drop-off. Across the 96-robot run the cluster collected 9,834 rounds of execution experience; 22 robots suffered 30 faults, and fault-caused machine-time loss was 1.3% of total runtime. On the algorithm side, FERT (failure-guided experience retrieval fine-tuning) reuses both successful trajectories and failures — successes supply learnable actions, failures indicate where to direct learning — addressing sparse reward on long-horizon tasks.

The framing shift is deliberate: humans move from supplying demonstrations to judging outcomes. CEL is explicitly positioned against HG-DAgger, the human-demonstration approach whose ceiling is set by how many skilled operators exist.

Why it matters. Embodied intelligence has had no equivalent of a scaling curve — no plot showing that adding something predictably improves task success. This is the first, and the input is the thing the industry already has a lot of: deployed robots. If it replicates, it inverts the current bottleneck. The standard story is that data is scarce because human demonstration is scarce; CEL’s claim is that fleet size converts directly into capability, which would make deployment itself the data engine and give a compounding advantage to whoever has the most machines in the field. Three cautions. (a) Baselines conflict across outlets: AgiBot’s own event material put phone packaging at 50.8% → 98.3%, the Oct 30 academic release says 79.2% → 98.3%; aggregate figures likewise appear as both 97.3% and 97.5%. Both are company/team-reported and neither has been reproduced. (b) The comparison set is six teleoperators on five tasks — real, but narrow. (c) The 1.3% fault-loss figure is genuinely encouraging for fleet operations, but it is measured over one campaign, not over months. Watch for independent replication and for whether FATS/FERT are released; the paradigm’s value depends on other labs being able to run it.

6. Optimus gets its first job: plugging the Supercharger into Cybercabs

What happened. Tesla executives confirmed in a Munro Live interview, reported by 财联社 / 21世纪经济报道 on October 30, that Optimus’s first real job is inserting and removing the Supercharger connector for Cybercab robotaxis — a transitional step pending full deployment of a wireless charging system.

The logic is a single missing half-metre. Cybercab already autonomously drives into a Supercharger stall and parks. Footage from an Austin Supercharger station shows the car waiting for a worker to lift the heavy cable, align it with the NACS port behind the rear bumper, and later return it — and to reposition the car if parking is off. That manual step is the one break in an otherwise fully automated operating loop, and it forces continued on-site staffing, which is precisely the cost robotaxi economics are supposed to eliminate. Optimus is scoped to close it: navigate to the charger, take the gun, plug in, unplug, return the gun, no human present.

Context from Tesla’s Q3 2026 call (Oct 21–22): Optimus entered “early production” on high-volume lines at Fremont, with the first “production-quality” units built; Musk described 20,000 units/week as “at least a year, maybe a little longer,” external company deliveries “around the end of next year,” and conceded Tesla is “building robots that do not have perfected hands.” Separately, the San Francisco Standard reported Oct 30 — citing robotics investor Modar Alaoui, unconfirmed by Tesla — that customer reservations may open around the Humanoids Summit, December 1–2.

The market read it as scenario expansion. By Oct 28 the 中证机器人指数 had risen 9.51% since mid-October; PE-TTM stood near 47–48×, against a five-year high of 83.83×. Chinese brokers cited as an additional driver that GPT-6 Astra can directly output robot end-effector poses — not just task decomposition — ranking first in a third-party AA-Robotics zero-shot comparison with 95% real-machine grasp success, which 华泰证券 called the first quantifiable validation of general-model capability migrating into the physical world.

Why it matters. Every humanoid company has shipped demos; almost none has published a job. This is the closest thing to a job specification yet released for Optimus, and its value is that it is unglamorous, bounded, economically legible and inside a business Tesla already operates. It also reframes the robotaxi story: the blocker on fully unmanned fleets is not the driving, it’s the last physical half-metre — exactly the class of problem humanoids are supposed to solve. Keep the caveats in view. The job is announced, not deployed: no timeline, no unit count, no success rate, and it is explicitly an interim measure before wireless charging makes it unnecessary. Musk’s own framing — unperfected hands, external delivery end of 2027 — is inconsistent with near-term volume. The GPT-6 Astra claim is a Chinese broker’s reading of a September OpenAI result relayed on Oct 30; the 95% figure is third-party and I have not seen the underlying eval. And the December reservation story is single-sourced and denied no comment by Tesla.

7. UBTech: 13,361 U1 orders, 1,500–2,000 units deliverable — and a question about who is buying

What happened. A October 30 analysis (海报新闻 / 大众网) of UBTech (9880.HK) puts numbers on the consumer humanoid gap.

The order/delivery gap. At its June 30 launch, UBTech announced 13,361 omnichannel orders for the consumer full-size humanoid U1 — more than twelve times the 1,079 full-size units sold in all of 2025. On its late-August interim call, management cut full-year U1 delivery guidance from “striving for over 10,000” to 1,500–2,000 — a reduction of more than 80%. JPMorgan downgraded from Overweight to Neutral and cut its target from HK$118 to HK$74. UBTech itself issued a share-price volatility notice in October flagging a trailing PE of 465×.

Why capacity fell short. The U1 head packs 19 DoF and two to three thousand components into a small volume to produce 30+ micro-expressions; biomimetic silicone skin yield is low; eyebrows and eyelashes are still implanted by hand; the micro servo is non-standard with an unproven supply chain; the line had to be built from scratch with no reuse of existing lines. The Guangxi plant’s start was delayed by more than two months. Founder Zhou Jian conceded it was “not fully anticipated.” Separately, the robot is positioned as a home companion, but the first three delivery customers — Shanghai Yongda, Beijing Boshi, and Korea’s Galaxy robot park — are all commercial venues doing guide and footfall work, and the Boshi arrangement is “retail and leasing in parallel,” meaning some units may be rented for display rather than sold.

The financials. H1 2026 revenue ¥1.269B, +104.2%, full-size humanoid revenue +1,445%; net loss ¥339M (narrowing ~23%); cash fell ¥4.888B → ¥2.326B, roughly halved in six months; operating cash flow negative. Gross margin improved past 75%, with Walker S2 BOM cost down from ~¥350k to ~¥180k. Management’s first explicit anchor: EBITDA positive in Q4 2026, net profit in Q4 2027. Receivables rose to ¥1.680B (+29%), now exceeding H1 revenue; inventory rose from ¥577M to ¥985M (+71%). Walker S2 delivered about 600 units in H1 against a full-year capacity plan pointing at the ten-thousand level.

The composition question. H1 new orders exceeded ¥1B, including Airbus’s first purchase of 100 Walker S2 units and some industrial repeat orders. But a material share of large B-end revenue comes from local government science-and-demonstration projects and robot data-collection centres — Jiujiang, Jiangxi committed ¥143M of fiscal money to buy 100+ Walker S2 units “to record the motions.” That budget is阶段性 (phase-limited).

Why it matters. This is the cleanest available test of whether consumer humanoids are a product category or a funding narrative, and the answer so far is the latter — but the more important finding is the revenue composition. UBTech grew 104% and lost ¥339M while selling substantially into government demonstration and data-collection budgets. That is a real business with a real expiry date on a chunk of its demand. The tell is JPMorgan’s apparently contradictory position: raising its 2030 global humanoid shipment forecast to 1.75 million units while downgrading the OEMs. Both can be right — a sector can grow tenfold without the current whole-body manufacturers capturing the margin, exactly the bifurcation visible with Unitree (profitable, ¥387M nine-month net profit, spending 19% of revenue on R&D) versus UBTech (still targeting Q4 2026 EBITDA breakeven). For anyone modelling this sector, the two questions to ask of every vendor are: how much of the order book is recurring commercial demand, and what is the unit-to-delivery conversion rate.

8. Bosch-backed 博银合创 puts an embodied robot on a real automotive-electronics mass-production line

What happened. 《科创板日报》 reported on October 30 that 博银合创 — incubated jointly by Bosch’s 博原资本 and 银河通用 (Galbot) — has opened its Suzhou headquarters and placed its industrial embodied robot into United Automotive Electronics (联电)’s Shanghai plant, targeting the flexible-handling gap at the end of automotive-electronics production lines.

The切口 is narrow and chosen well. Plant manager Lu Ni told the paper that welding, tightening and dispensing are already fully automated, but line-side material handling, loading/unloading and loose-part sorting remain heavily manual: “Traditional industrial robots demand high positioning accuracy; AGVs solved plant-wide transport, but the ’last kilometre’ of precise pick-and-place has long lacked a mature automated solution.” Commissioning began in June; the first robot entered the mass-production line this month on a standardised pick-and-place operation. Current accuracy 80–90%, hot-swappable battery with roughly 6 hours per charge, still requiring occasional manual correction of placement deviation, and cycle-time efficiency has not yet reached a skilled worker.

The vertical strategy is the substance. CEO Ren Jingwei says its in-house Bolt model deliberately forgoes full-scene generality: the company built a real-machine data collection centre at Suzhou that teleoperates replicas of actual production stations and captures full sequences with force feedback, on the argument that teleop-collected native station data trains far better-adapted models than simulation data. The pitch is that first-scene commissioning costs about the same as traditional robotics, but marginal cost falls as the model is reused across similar stations — matching automotive’s fast iteration cadence. Galbot’s AstraBrain model and AstraData infrastructure supply the base; Galbot targets general service and retail while 博银合创 takes industrial verticals. Suzhou targets 4,000 units/year by 2027, with a long-range plan of ten-thousand-level annual capacity; priority sectors are automotive parts, 3C and new energy.

The industry-wide constraint is stated plainly. Three bottlenecks: cycle efficiency, cost model, scene generalisation. The sector’s rule of thumb is that an industrial embodied robot must fall below ¥200,000 whole-unit cost to be economically viable for large-scale labour substitution. Ren’s counter-argument is that industrial value shouldn’t be benchmarked only against human cycle time: robots work 7×24 on the night shifts and high-intensity repetitive posts that workers avoid, so the综合 input-output ratio can already work. Investors quoted by the paper call the 联电 trial a representative case of Chinese industrial embodied intelligence moving from lab prototype to real mass-production scene — while noting the vertical route deploys faster but has a narrower applicability boundary.

Why it matters. Most “robot enters factory” stories are pilots in a showcase bay. This one is on a live automotive-electronics line at a tier-one supplier, with the unflattering numbers attached: 80–90% accuracy, cycle time below a human, cost above the ¥200k threshold. That triad is the honest state of the art, and it maps exactly onto the three bottlenecks the sector names for itself. The strategic content is the data thesis — teleoperation on replicated real stations with force feedback beats simulation — which is the same claim being made from the opposite direction by CEL (item 5): one says get data from deployed peers, the other says get it from real stations rather than sim. Both are bets that the binding input is native physical data, not model scale. If both hold, the moat in industrial embodied AI is access to production lines, which favours incumbents with factory relationships over model-first startups.

Radar — also worth knowing

Six trends this window

1. Google conceded the agent layer and bet on the surface. Shutting Jules and opening Android Studio to 40+ third-party agents via Agent Client Protocol in the same 48 hours is not two strategies, it’s one: the IDE, the plugin registry, the mobile toolchain and the protocol are the assets; the agent is interchangeable. Antigravity’s Plugin Marketplace — Skills + MCP servers + Rules + Hooks + Custom Agents as a one-click installable bundle — is where Google expects the lock-in to live. Every other lab is making the opposite bet on the harness itself.

2. Agent convenience is now the attack surface, and it is being catalogued systematically. EMBRACE is the third consecutive day of this pattern: a preview browser, a web worker outside the folder check, a sandbox the agent turns off by itself. The recurring shape is that every capability granted to an agent is a boundary it can route around. Two of five tools were unpatched at publication, and the defensible position is the one Forever states plainly — assume the agent will attempt to bypass its own safeguards.

3. Agent cost engineering became measurable, and the waste is structural. TokenEconomizer’s profile is the finding: 16 of 23 loaded tool schemas never called, 3 covering 98% of calls; 30% of calls are repeats after five tasks; a third of spend going to runs that will fail. Add Factory’s PTC (Oct 29) and Microsoft’s harness opening (Oct 29) and the picture is consistent — the harness, not the model, is where cost and accuracy are won, and it is now being measured rather than asserted. Expect a token-efficiency layer to become standard middleware.

4. Embodied intelligence got its first scaling law, and the input is fleet size. CEL’s 16→96 robots at constant training time (86.7% → 98.3%) is the first curve in this field whose axis is robots rather than parameters. If it replicates it inverts the data-scarcity narrative: deployment becomes the data engine, and 博银合创’s insistence that teleop-on-real-stations beats simulation is the same thesis from the factory side. Both predict that the moat is access to physical work, not model scale.

5. Physical AI’s first real jobs are unglamorous, and that is the signal. Optimus plugging a Supercharger into a Cybercab; 博银合创 doing pick-and-place at the end of an automotive-electronics line; AgiBot’s convenience-store shift scheduled for Nov 2–3. None is a demo. All are narrow, bounded, economically legible, and inside businesses that already exist. The comparable honesty is in the failure numbers: 80–90% accuracy, cycle time below a skilled human, cost above the ¥200k threshold.

6. The embodied ledger’s question shifted from “can you build it” to “who pays, and does it recur.” UBTech: 104% revenue growth, ¥339M loss, half its cash gone in six months, 13,361 orders against 1,500–2,000 deliveries, and a material share of revenue from government demonstration and data-collection budgets. JPMorgan raising its 2030 shipment forecast to 1.75M units while downgrading the OEMs is the whole thesis in one line — the sector compounds; the current manufacturers may not. Unitree’s profitability against UBTech’s Q4-2026 EBITDA target is the same split.

Sources & method

• AI HOT agent/latest feeds were unreachable this run (endpoint 404); the entire issue was assembled from search plus direct fetches — the eighth consecutive window with no robotics items available from that feed.
• Direct fetches: helpnetsecurity.com/2026/10/29/research-ai-assistant-browser-vulnerabilities (EMBRACE, per-vendor status); 163.com 海报新闻/大众网 (UBTech U1 orders, delivery guidance, H1 financials, Jiujiang data centre); sohu.com 科创板日报 Oct 30 (博银合创 × 联电, Bolt model, ¥200k threshold); hqwc.cn/a/1157930.html (geek-daily roundup — dated Aug 15 in its own header despite an Oct 30 publish stamp, so its items were excluded from the body).
• Google (item 1): himiao.ai, pivotnews.ai, timesofai.com, nvdium.com (Jules timeline, export scope, Plugin Marketplace partners); themorningbuild.com digest Oct 30 (Android Studio BYOA, Agent Client Protocol, Gemini Nano 4 Agentic API, MCP/AppFunctions/A2UI, Android Bench 2.0).
• TokenEconomizer (item 3): digestai.news and MarkTechPost (Oct 30), Microsoft Research × UC Berkeley; profiling figures and per-repository results as reported.
• OpenAI Private Intelligence (item 4): agipulse.org item 1074 (Oct 31, translating OpenAI’s Oct 30 X post and product page), plus the Aug 19 / Sept 22 / Sept 29 / Oct 26 timeline.
• CEL (item 5): 光明网, 央广网, 中国证券网 (cnstock 820391), 封面新闻 (all Oct 30); cross-checked against AgiBot’s own Oct 29 Innovation Day material.
• Optimus (item 6): 21世纪经济报道/财联社 via Toutiao (Oct 30), 163.com analysis, thenovtech.com (Q3 call transcript detail), basenor.com/SF Standard (Dec reservations), 每日经济新闻 (index/PE data).
• Radar: havoptic.com and claudenews.com.br (release notes), llmboard.ai (OSWorld 2.0), yingzheng.com (Mistral FrontierCode), nocode.mba, rainvent.ai, agipulse.org 1076 (AIPREF), physixfrontier.com, Sifted/Humanoids Daily via 全球机器人情报日刊, 中国经营报, 科创板日报, toutiao AI roundup Oct 30.

Quality check

• Jules/Antigravity: shutdown dates and export scope are from Google’s changelog and X post as relayed by four outlets; the absence of a migration path, a background-task date and a free-tier equivalent is reported consistently but is an absence, not a statement. The Plugin Marketplace partner list varies slightly by outlet (nvdium adds Canva/CrowdStrike/monday.com/Unity/Ping Identity/SentinelOne to himiao’s Atlassian/Figma/Stripe). Android Studio BYOA (“40+ agents”) is from The Decoder via Morning Build, not a Google blog I read directly.
• EMBRACE: one research firm’s account relayed by a news site. Vendor responses are characterised (“declined, citing Workspace Trust”) but no vendor advisory or statement was published in the article, so per-vendor patch status should be treated as unverified. The Cursor sandbox self-disable is model-dependent and, in Forever’s test, launched Calculator — the escalation to arbitrary code is the researchers’ inference, not an observed outcome.
• TokenEconomizer: all figures are the authors’ own, on 277 SWE-Bench tasks across 5 settings; the Pass@1 −1.0 point regression on Opus 4.8 is theirs to report and I have kept it. The 6.4–9.3% result on ReposVul shows the method does not generalise uniformly. Comparison with Bito Governor’s claimed 70% is across different task sets and should not be read as a ranking.
• OpenAI Private Intelligence: private inference is API preview with eligibility gating; no pricing, model list, hardware or GA date. The whitepaper excludes Bedrock and other third-party channels. Not previously covered.
• CEL: baseline conflict is real and unresolved — phone packaging appears as 50.8% → 98.3% in AgiBot’s Oct 29 event material and 79.2% → 98.3% in the Oct 30 academic release; aggregate success appears as both 97.3% and 97.5%. All figures are team-reported with no independent reproduction. The human comparison is six teleoperators across five tasks. The 1.3% fault-time-loss figure is from a single campaign. This item extends yesterday’s AgiBot Innovation Day coverage — the scaling curve, FATS/RLinf detail, 9,834 rounds and the 16→96 comparison are new; Worken, GO-3, 灵渠 OS and the Nov 2–3 store run are not repeated here.
• Optimus: the charging-gun job is announced, not deployed — no timeline, unit count or success rate, and it is explicitly interim before wireless charging. The December reservation story is single-sourced to one investor and unacknowledged by Tesla. The GPT-6 Astra end-effector/95% claim is a Chinese broker citing a September third-party eval; I have not seen the underlying evaluation. Index and PE figures are as of Oct 28–30.
• UBTech: H1 figures per the company; delivery guidance cut is from the late-August interim call; JPMorgan’s downgrade and target are sell-side opinion. The characterisation of government demonstration/data-centre procurement as phase-limited is the outlet’s analysis, not a company statement. Unitree comparison figures are from its listing documents (already covered Oct 29–30).
• 博银合创: accuracy, battery life, commissioning dates and capacity plans are company statements to 科创板日报; the ¥200k threshold is the report’s summary of “industry-wide” estimates, not a disclosed cost for this robot. The three bottlenecks are the paper’s framing.
• Dedup: Factory Droid 1.0, AgiBot Innovation Day (Worken/GO-3/灵渠 OS/Nov 2–3 run), OpenAI’s L0–L3 monitoring stack, Mistral Large 4 weights, Wiz agentic-attacks, Xinhua/CAICT report, Unitree Q3 and Proton (all Oct 29–30 items in the previous issue) are not repeated as body items; Mistral appears only as a FrontierCode delta, Unitree only as a comparator inside item 7, Proton only in the trends. Excluded as previously covered or out of window: Kimi K3.1 Pro launch (Oct 28), GitHub Octoverse/Universe, Gemini 4 Carbon, Anthropic S-1, Bito Governor, RoboParty RP1, XPeng IRON SOP timeline.