EAIDaily — October 25, 2026

Daily intelligence brief on AI coding and embodied intelligence Coverage window: October 24, 2026 08:15 → October 25, 2026 08:15 (Beijing time) Focus: AI coding / agentic software engineering · embodied intelligence / humanoid robotics

Today’s throughline: the runtime layer underneath agents got built in public, by two hyperscalers, in 24 hours. Google published an account of Gemini 4 Argon running unattended for 11 hours against Google’s own 200,000-line Agent Development Kit — it landed two accepted optimizations and had four rejected, with Google naming review, debugging and testing as the real bottleneck. Google also unveiled Agent Compute Engine, where each agent gets a hardware-isolated microVM that never holds a credential (real ones get swapped in at network egress by a gateway sitting outside the sandbox), and open-sourced Agent Substrate on GKE to match. AWS answered the next day with Infinity, which suspends an agent at every tool call so idle agents cost nothing — about 100 KB each. In the physical world the opposite of showmanship happened twice: AgiBot put 110 mass-production robots on a stage in Jinan for a zero-failure Guinness record, and Shanghai wrapped a competition where employers set the exam, the state released a national-standard evaluation ruler for embodied AI, and the failures were reported as loudly as the wins.

1. Gemini 4 Argon spent 11 hours optimizing Google’s own agent framework — and got 4 of 6 changes rejected

What happened. Google engineers pointed Gemini 4 Argon at the Agent Development Kit (ADK) itself — the open-source Python framework Google uses to build agents — and let it run unattended for 11 hours. The target was deliberately hard: over 200,000 lines of code across roughly 800 files, with nearly 17,000 unit tests. The task was also deliberately hard: not “build me an app in an empty folder,” but performance work on a mature codebase, where the slowdown is almost never in the file where the symptom shows up.

The setup is the most transferable part of the report. Google first ran a short trial to discover what the prompt needed to contain for the agent to follow team conventions, then encoded those as repository-specific guidelines in .agents/skills files — described as “like an onboarding guide for a new engineer” — and told the agent to load only the single skill required for the task, to keep the context window clean. The final instruction used Antigravity’s /goal mode so the objective survived context compaction:

Make ADK add no noticeable overhead to any agent: at startup, in long sessions, in multi-agent and workflow setups. Keep what the model receives and the public API unchanged, and back each change with a test that would catch a regression. Open one pull request against main that the ADK maintainers would approve, with before and after numbers in its description. Don’t merge it.

What came out. Argon’s logs showed it reading the guidelines, establishing performance baselines, proposing optimizations, reviewing its own changes, and running all 16,892 open-source unit tests before finishing. Two changes were accepted:

Google’s argument for why this matters is not the headline percentages — it is compounding. A human never notices 28 ms or 78 ms on a single LLM call that takes seconds. But across 158 internal tool-calling turns, removing those copies cut the full run from 65.6 s to 56.1 s — roughly ten seconds of pure framework overhead, i.e. ten seconds the model’s own latency is now the only thing you are waiting on.

The part everyone should read twice. Argon also produced four modifications that were rejected in developer review, for concerns about reliability, maintainability, compatibility, or insufficient measurable benefit relative to added complexity. That is a 33% hit rate, and Google says the quiet part out loud: agent effectiveness “remains currently limited by the time and effort required for post-coding tasks, including PR review, debugging, and testing.”

Why it matters. For two years the scoreboard was can it write code. This is one of the first credible vendor accounts of an agent doing the thing engineers actually respect — cross-module performance work on a brownfield repository — and it immediately relocates the constraint to acceptance. Google credits Argon with doing what a profiler cannot: tracing every caller, mock target and shared reference across ~40 files so the faster path doesn’t silently break dependent applications. Note the caveats: this is Google on its own codebase, one reported run, self-reported; Gemini 4 Argon’s context window, price and benchmark scores were not disclosed. (Sources: InfoQ, Sergio De Simone, Oct 24; Google engineering write-up.)

2. AWS open-sources Infinity: treat every tool call like an await, and 75,000 idle agents fit in 8 GB

What happened. On October 23, AWS released Infinity under Apache 2.0 — part of Hydro, its collaboration with UC Berkeley, written in Rust, blogged by Shadaj Laddad as an explicit early research release. The premise is that agents spend most of their lives blocked: a coding agent waits 20 minutes on a build, a triage agent waits hours for the next pull request, an assistant waits overnight for an approval — and in most harnesses the process and its memory sit there, unable to respond to anything else.

The trick is borrowed from programming languages: an agent can be suspended exactly like a stackless coroutine. At the moment a tool is called, the conversation history already contains everything needed to continue, so Infinity dispatches the call, persists history, and drops all in-memory state. Each turn runs as a bounded slice — load history, one model completion, persist output, dispatch at most one tool call, end. Nothing ever blocks inside a slice; every resumption arrives as a message on a shared input queue, with per-thread ordering guarantees and cross-thread parallelism.

The numbers. ~100 KB of resident memory per idle agent, which AWS translates into ~75,000 concurrent agents in 8 GB — the memory of a Raspberry Pi 5. Their benchmark ran 80,000 agents × 20 turns = 3.2 million completions on a single thread of an AMD EPYC 9R14. Important caveat: the benchmark used a simulated model and in-memory stores. The Pi is a memory-size comparison; the test did not run on one.

The deployment story is why this is more than a parlor trick. Because suspended agents are pure data, serverless functions can exit while tools run: AWS’s reference architecture uses SQS FIFO as the input queue, Lambda per step, Aurora DSQL for conversation histories, DynamoDB for processed-message tracking, SQS delays plus EventBridge Scheduler for timers, with CDK templates included. Laddad’s line: an agent waiting on a three-day CI pipeline uses no compute until the pipeline reports its result. A coding agent, Infinity Code, is built on the runtime, and the repo describes a Reactive Agent Protocol claiming existing MCP servers run unchanged through a compatibility layer.

The genuinely novel bit is synthetic tool calls. Model APIs expect exactly one result per tool call, leaving no clean way to tell an agent something new happened. Infinity injects a tool call the model never made (receive_event__injected) whose result carries the event — and uses the same mechanism for agent-to-agent work (spawn_thread, report_to_parent), keeping child results out of the parent’s context window. Laddad lists the open questions himself: how models trained on step-by-step tool use cope with interleaved events, scheduling priorities, compaction across agent trees, and how to verify an incoming event is authentic.

Why it matters. Every hyperscaler is currently repricing what an agent fleet costs, and they attacked two different halves of the same curve this week: Google solved the cost of isolation and credential trust; AWS solved the cost of waiting. Together they say the same thing — agent economics are now decided by the runtime, not the model. Start paying attention to accept/reject rates and idle-hour costs the way you once watched tokens-per-dollar.

3. Google Cloud’s Agent Compute Engine: one microVM per agent, and the credentials live outside the sandbox

What happened. On October 22 at AGNTCon + MCPCon in San Jose, Google Cloud announced Agent Compute Engine (ACE), which already runs Gemini, Antigravity and Google’s evaluation and RL pipelines internally. Managed preview is “coming months”; pricing is unpublished. The design splits the agent in two: a harness outside the execution environment that calls the model and chooses actions, and a sandbox — one hardware-isolated microVM per agent session, with its own guest kernel and root access, running on hosts with Google’s Titanium IPU and Titan security silicon.

The security claim is unusually strong and unusually concrete:

Assume an attacker controls everything inside the sandbox, and make sure they find nothing there to use.

No OAuth refresh tokens, no API keys, no service-account keys inside the guest, and no access to the instance metadata server. Instead each microVM is handed an ephemeral, session-scoped SPIFFE identity certificate plus a short-lived token that proves which agent this is but carries no backend credential. Outbound traffic crosses a point-to-point virtio-vsock boundary to an unprivileged local proxy and then to an external Agent Gateway, which validates three things — user delegation token, agent SPIFFE identity, and Titanium-attested node identity — and only then performs an in-flight token swap, inserting the real backend credential onto the wire on its way out. Attestation keys live in Titan/Titanium silicon, outside the guest; block-storage attachment is hardware-mediated, so a compromised sandbox cannot mount another tenant’s volume.

This answers the canonical enterprise failure: an agent with a repo token reads a poisoned issue, gets told to exfiltrate keys — and finds no keys. It can still misuse the access it legitimately has for that session; it just cannot escalate or walk away with anything reusable. Google also separates four identities (user, delegated on-behalf-of grant, agent workload identity, optional workforce identity), giving the agent a first-class cryptographic identity that is neither a person nor a service.

On cost, Google made the same bet AWS did — via a different mechanism. MicroVMs boot in ~125 ms at ~5 MiB of private per-agent memory because hundreds of agents share one OS copy through copy-on-write + Direct Access. Idle agents are checkpointed and restored in under 500 ms, releasing CPU and memory to active agents, with per-agent Hyperdisk keeping installed libraries and workspaces between sessions. Google also published primitives that matter directly to coding agents: clone, fork, rewind and rollback, so you can branch a workspace into parallel experiments without rebuilding environments.

Alongside it, Agent Substrate on GKE ships now as open source — the same model on any Kubernetes cluster (gVisor or Cloud Hypervisor), with Google claiming up to 10× higher compute density and a 200,000-agent validation run sustaining 500 activations/second at sub-500 ms suspend/resume. Google also extended its governance stack — Agent Gateway, Agent Registry, Agent Identity — beyond Google Cloud to multi-cloud, on-prem and managed endpoints.

Why it matters. Read items 2 and 3 together: the two hyperscalers independently concluded that generic cloud compute is the wrong substrate for agents, and shipped. Also note the sequence — this lands four days after the LiteLLM supply-chain compromise (Oct 21) stole API keys out of agent installers, and two days after OpenAI published containment architecture for its own training cluster. “No static credentials inside the runtime” is rapidly becoming a procurement checkbox rather than a research position.

4. Anthropic stops chatting: Dashboards and Motion arrive while Docs, Slides and Design leave beta

What happened. Anthropic shipped three things in one announcement: Dashboards (live, connected to your data), Motion (animation), and the removal of the beta label from Docs, Slides and Design — now available on all plans including Free. Since those three landed inside every Claude conversation, users have produced more than 45 million documents, presentations and designs. The two new ones are gated: Dashboards is beta and paid-plans only, Motion is beta and Team/Enterprise only.

What Dashboards actually does is worth spelling out because it collapses a workflow most companies have simply written off. Connect Claude to a warehouse (BigQuery, Snowflake, Databricks) or a CRM like Salesforce, ask in plain language what happened to signups this week versus last month, and Claude writes the query, runs it, and builds the dashboard — which then tracks the data as it changes. Every number is clickable to reveal the query behind it, Claude can explain that query on request, and each chart is stamped with when its data was last refreshed. That last detail is the difference between a chart and a decision.

Why it matters. This is the artifact race, and it is moving fast. OpenAI launched ChatGPT Space at DevDay; under ten days later Claude shipped dashboards and motion, and — per reactions Anthropic’s own posts attracted — users immediately started asking not for more surface but for a Dashboards API so they can generate live boards inside their own applications. The expectation being expressed is that the assistant’s job ends in a deliverable, not a suggestion. Pair this with Claude’s auto mode becoming the default on web, desktop and mobile for Pro/Max/Team (automated checks replacing step-by-step approval, manual switch retained), and the direction is consistent: the agentic path is becoming the default path, with the controls moving from per-action approval to policy-level guardrails. Note the honest caveat — product-usage figures (45M artifacts) are Anthropic’s own and unverified.

5. The review-and-cost toolchain for coding agents is being written in public

While Google was naming review and testing as the bottleneck, four small open-source projects shipped aimed at exactly that seam. All are author-reported and small-sample, but together they map where practitioners actually hurt.

Why it matters. Every one of these targets the after — reviewing what the agent did, proving it described all of it, and paying less for the privilege. task-view “every changed line must be accounted for” is arguably the single most useful primitive to appear this month for teams letting agents touch production code. None of these numbers is independently verified and the samples are tiny (18 bug fixes), so treat them as leads, not benchmarks.

6. AgiBot put 110 mass-production humanoids on one stage, and the interesting claim is that none of them were special

What happened. On October 24 at the Jinan Yellow River International Convention & Exhibition Center, 110 AgiBot 远征 A3 full-size humanoids performed a synchronized routine for 5 minutes 43 seconds with zero errors, certified on stage by Guinness World Records adjudicator Hou Ying as the record for the most humanoids dancing simultaneously. AgiBot partner and SVP Wang Chuang was explicit about the reason it matters:

Every robot on stage today is a mass-production unit. Not one is a specially tuned demo machine. Joint zero positions, torque curves and action libraries are uniform at the line level — and that consistency lets 110 robots behave like one.

That framing inverts the usual logic of these spectacles. A lab demo needs one success out of a hundred tries; a production line needs zero failures out of a hundred units, because one failure is a safety incident. The standard itself was chosen accordingly: not “how well can it dance,” but whether a hundred-plus robots can go error-free for five minutes.

The technical body of work underneath is fleet software, not choreography. The robots ran autonomously end to end — stepping out of their shipping crates, self-calibrating into formation positions, then executing — with no per-robot handheld intervention. The orchestrator is LinkBots, AgiBot’s in-house multi-robot control system, whose three capabilities map almost one-to-one onto warehouse requirements: real-time formation changes computed from live state rather than pre-stored playback; autonomous navigation and walking with UWB positioning accurate to under 10 cm; and per-formation tasking under one orchestrator so groups don’t interfere with each other. Wang calls that trio “the technical watershed between a robot being a demo piece and being productive.”

Industrialization is stated in units, not aspirations: the Jinan base is planning a fully automated line, expected to be visible around mid-next-year, with a single line capable of 8,000 units/year and four lines in the long-term plan; reducers and other core parts come from deeply-bound suppliers such as Henggong Precision. On market position, Chinese coverage of IDC data puts H1 2026 global humanoid shipments near 25,000 units (+432.1% YoY) with AgiBot shipping >8,600 units in H1 — roughly 35% globally and >45% in China. These robots are already doing unglamorous work: guidance, passenger-flow support, security patrol and cultural interpretation in Jinan Olympic Sports Center and Zhuanshan West Road metro stations.

Why it matters. The competitive claim has changed category. It used to be “our robot can do X.” It is now “we can make a thousand identical robots that all do X without failing, and supply the parts.” That is a manufacturing-and-supply-chain moat with a software control layer on top, and it is much harder to copy than a demo. Caveat: the shipment figures are IDC data as relayed by Chinese media and AgiBot’s own statements — no independent audit.

7. Shanghai’s embodied-intelligence exam wrapped — and China shipped the ruler it will be graded with

What happened. The 2026 Global Developer Pioneer Conference (GDPS) and International Embodied Intelligence Skills Competition ran October 23–25 at Zhangjiang Science Hall in Shanghai, closing today: 800-plus teams and roughly 1,000 humanoid robots, across 9 tracks and 23 events (framed at the opening ceremony as ten real-scene tracks) spanning energy inspection, quality inspection, supermarket retail, emergency rescue, cultural-tourism guiding and healthcare. The tracks were distilled from more than a thousand real embodied-intelligence scenarios solicited nationwide by MIIT, with energy inspection the most popular and logistics loading/stacking drawing 150+ teams.

The organizing principle was “industry sets the exam, technology answers it” — with a 2,000 m² real-scene zone built to genuine specification. Bailian Group reconstructed a Lianhua supermarket down to real aisle widths and shelf heights, with self-checkout scanners in place, and reduced a store associate’s day to five exam questions: receiving, restocking, shelf straightening, tray service rounds, and picking/packing. There is a 1:1 replica substation and simulated earthquake rubble. The Shanghai Municipal Quality Inspection institute’s quality track requires a continuous six-step chain — receive and sample a milk tablet, weigh and prepare, load a centrifuge, retrieve and return the sample, check state, then sort by verdict — using a single end-effector for grasping, screwing and placing. Top-three finishers per track get matched with the scenario owner, which converts a trophy into a plausible pilot line.

The most consequential thing released happened off the competition floor. Fan Kefeng, vice president of the China Electronics Standardization Institute, released an embodied-intelligence evaluation benchmark system and matching evaluation tools, built on the existing series of national standards — an attempt at one authoritative measurement ruler for a technology that has been iterating faster than anyone can measure it. Alongside it, vice minister-level remarks framed the policy stance plainly (MIIT Science and Technology Department deputy director Gan Xiaobin: practical skills driven by demand, using competition to accelerate deployment, open collaboration for the ecosystem), and Shanghai put numbers behind the scene-building: AI industry scale above ¥330 billion in H1 2026, +22.6% YoY.

The reporting included the failures, which is the tell. Xinhua’s dispatch describes a team repeatedly failing at magnetic shelf price tags — they must be brought very close to the shelf to catch, and that gap is too small for the gripper — with the team admitting they had trained mostly on loading and picking and that “seemingly simple” real requirements like carrying a tray or swapping a price label are exactly the questions robots must answer. Yang Xindi, CTO of Shanghai JuShen Huizhi, makes the same point at the structural level: no single action in the QC chain is hard; chaining them while handling randomized orientations is, and the real test is whether it runs stably for a long time and behaves correctly when something goes wrong.

Why it matters. This is verification being built as public infrastructure: the employer authors the task, live instrumentation produces efficiency/stability/ROI numbers, a national-standards body supplies the measurement ruler, and the winners get a customer. It also means the narrative is no longer controlled by whoever edits the highlight reel — which is what makes a 5-minute Guangxi fault, a dropped sample, or a magnetic label nobody can grip genuinely informative. Note: we flagged this event two days ago as “no results yet”; the competition concludes today, so per-track winners are still landing, and the benchmark system’s actual contents have not been published in detail.

8. Brains and bodies decouple: Unitree builds the body for Galaxy General’s “native agentic robot,” and AgiBot opens its reasoning data

What happened — the joint product. At the 2026 World Physical AI Conference and Unitree partner ecosystem conference, Galaxy General (银河通用) and Unitree announced that Galbot ET1 (银河星仔) — Galaxy General’s first bipedal robot, first shown at the World Robot Conference this past August — was jointly developed by the two companies. The announcement frames it as the world’s first “native agentic robot”: rather than running preset scripts for performances, ET1 supports real-time interaction, generating speech and motion on the fly for improvised response and human-robot perception, driven by Galaxy’s “Galaxy Star Brain” which the company says was trained on 100,000 hours of human motion data, letting it identify and reproduce human motion features from observation.

The division of labor is the interesting part: Unitree on the hardware body and production polish, Galaxy General on the embodied brain — and their statement describes deep coordination across product definition, hardware refinement and embodied-algorithm capability, rather than a reseller relationship.

What happened — the open data. In parallel, AgiBot moved to daily releases as part of its “AI WAVE” seven-part sequence. Day one opened an “embodied reasoning” dataset — structured to link goal understanding, process memory and action execution so a robot can both think it through and do it — following the late-2024 open-sourcing of its million-episode real-robot AgiBot World dataset. This year’s four releases span imitation learning, diverse interaction, real-robot reinforcement learning, and embodied reasoning. The second item is Genie Studio 2.0, a one-stop full-process platform aimed at “actually putting robots to work”: Flywheel (model iteration), Agent (modular task orchestration), and Monitor (multi-robot fleet ops), covering the collect–train–test–deploy–orchestrate–manage lifecycle, with a claim of taking a brand-new scenario from zero to verified deployment in as little as 7 days. Partner and embodied-business president Yao Maoqing gives the motive plainly: “we want to empower the whole industry; only if the industry rises can we develop better.”

Why it matters. Two value-migration signals in one day. First, bodies are becoming a foundry service — the differentiated asset is the agent brain and the data that trains it, which is precisely the bet OpenAI made when Altman framed intelligence, not hardware, as the hard problem for its own humanoid program. Second, data and deployment tooling are being given away to make someone else’s robots better, because the prize is the industry-wide standard position rather than the margin on one robot. Treat “world’s first native agentic robot” and the 100,000-hour figure as company claims; no independent verification exists, and no shipping date or price was announced.

Radar — also worth watching

• Standards week recap (context; the Agent Skills donation itself we covered Oct 22–23). At the same AGNTCon + MCPCon event, AAIF took in four new projects in a single day — Agent Skills (Anthropic), Open Responses (OpenAI), OpenEnv (Hugging Face), OpenSandbox (Alibaba) — taking the Linux Foundation-directed foundation to 10 projects in its first year and 305 member organizations. Maintained by individual engineers across Anthropic, Google, OpenAI, led by David Soria Parra and Mahesh Murag. Alleged install bases (1M skills, 280M installs) remain vendor-reported with no stated counting method.
• China clears exports: DeepSeek and MiniMax approved for global model deployment (Chinese state media, citing Bloomberg), landing a day after DeepSeek’s reported **¥100B ($15B)** raise and days after the US DHS exclusion order. Separately, Apple Intelligence was approved in China, built with Alibaba’s Qwen and Baidu, rolling out live translation, writing tools and image cleanup now, with a wider launch targeted at iOS 27.3 in December — still without Siri AI, Image Playground or newer Genmoji.
• Credentials keep leaving the runtime. 1Password for Agents issues cryptographically signed short-lived session tokens bounded by least-privilege envelopes (endpoints, budget caps, time windows) with tamper-evident audit logs; Vercel’s CEO publicly called for “killing static credentials” and doing for OIDC what TLS did for encryption. Same thesis as Google’s ACE in item 3, applied at the password-manager layer.
• Chinese brokerages turn uniformly bullish on robotics. CITIC Securities argues liquidity, valuation, policy, technology and chain-leader catalysts all match the pattern of the previous seven robotics rallies; Zhejiang Securities sees bodies entering an accelerated volume phase (tens of thousands → hundreds of thousands of units) with upstream components monetizing first because bodies face price competition while standardized parts get multi-customer scale; Sinoline notes Chinese vendors swept H1 global humanoid revenue top-5.
• Consumer robots go to a lifestyle festival. Shangwei New Materials’ Qiyuan (启元) Robotics brought Q1 and T1 to Xiaohongshu’s first MANANA Tomorrow Park (Haikou, Oct 23–25); T1 switches between humanoid and quadruped forms for carrying and follow-cam. Symbolically useful: this is distribution to consumers, not to factory buyers.
• xAI’s partner network reached 130+ companies, with Grok Bot past 800,000 weekly paid users — reportedly more than doubling in a month.
• David Silver (Ineffable Intelligence) argues superintelligence arrives from systems that keep discovering through experience and RL, not from scaling human-data LLMs. Direct relevance to item 8’s real-robot RL datasets and to the “deploy to learn” logic behind embodied fleets.
• Open-source coding-agent scenery: task-view, Token Beggar, lcc and Slate Term all trended the same week (see item 5); GitHub’s AI-coding-agent topic now lists opencode at 220.6K stars, spec-kit 142.6K, codex 130.7K, reflecting how quickly the harness layer consolidated.

Key trends across today’s items

1. The agent runtime became a product category this week. Google published isolation-without-credentials and open-sourced a 10×-density Kubernetes substrate; AWS published suspend-at-tool-call and made idle agents free. Both concluded that general-purpose cloud compute is the wrong shape for agent workloads — which means the next round of platform competition is about sandbox economics and trust boundaries, not model quality.
2. Generation stopped being the bottleneck; acceptance is. Google’s own result is the cleanest data point available: 2 accepted, 4 rejected, with review/debug/test named as the constraint. Watch merge-acceptance rate and review cost over the next few quarters — they are replacing benchmark scores as the honest productivity metric.
3. “No static credentials inside the runtime” is becoming a default requirement. Three independent moves (Google ACE’s egress token swap, 1Password’s scoped delegation, Vercel’s OIDC push) converge on the same answer within days, in the immediate shadow of the LiteLLM supply-chain theft.
4. Embodied AI’s moat is now manufacturing consistency, not capability. 110 indistinguishable production robots running error-free for 5:43 is a supply-chain and fleet-software claim. The next differentiators are line-level uniformity, ~10 cm fleet localization, and 8,000-unit/year capacity.
5. Measurement became infrastructure. A national-standards-based evaluation benchmark plus employer-authored exams plus live ROI instrumentation means the industry can no longer select its own highlight reel — and the failures (magnetic labels, six-step QC chains) are being published.
6. Brains, bodies and data are separating into different companies. Unitree supplies bodies, Galaxy General supplies the agent brain, AgiBot open-sources the data and the deployment platform. Value accrues to whoever owns the learning loop and the toolchain; the body is trending toward commodity.

Sources

AI HOT curated feed (/api/v1/agent/latest, 3 items in the 24h window — zero robotics items again) plus six passes of verification across:

• Coding / runtime: InfoQ (Sergio De Simone) for Gemini 4 Argon × ADK; AWS Open Source Blog (Shadaj Laddad) for Infinity; Google Cloud engineering deep dive, Tech Times, Stack Futures, AppStack Insider, NeoTeo (via The New Stack) for Agent Compute Engine and Agent Substrate; Sohu/新智元 for the Claude Dashboards/Motion announcement; my2cents.ai and Show HN for the open-source tool cluster; Buttondown 每日AI洞察 for auto-mode and OpenResponses confirmation.
• Embodied: 科技日报 and 齐鲁壹点/网易 果然财经 for the AgiBot Guinness record, LinkBots and Jinan base; 新华社, 上海证券报/中国证券网, 证券时报, 东方财富 for GDPS, the ten real-scene tracks, the evaluation benchmark system and the honest failures; 环球网/科创板日报 for Galaxy General × Unitree; 腾讯新闻/太平洋科技 for the AI WAVE releases and Genie Studio 2.0.
• Context: aiproplaybook and techbytes daily roundups for cross-checking the Oct 23–24 vendor announcements.

Quality check

• Vendor-reported and flagged inline: all Argon/ADK numbers (Google’s own codebase, single unattended run, model specs undisclosed); all Agent Compute Engine figures including the 10× density claim, which belongs to the open-source Agent Substrate rather than the managed product, and 200,000-agent/500-activations-per-second from Google’s own validation; the 75,000-agents-in-8 GB figure comes from a benchmark using a simulated model on a single EPYC thread with in-memory stores — the Raspberry Pi 5 reference is a memory-size comparison, not a test platform.
• Unverified claims marked as such: “world’s first native agentic robot” and the 100,000-hour training corpus for Galaxy Star Brain are company statements with no independent check and no announced price or ship date; the 45 million Claude artifacts figure is Anthropic’s own and unaudited; open-source tool results (18 bug fixes for Token Beggar, 299 sessions for lcc) are author-reported and small-sample, with lcc’s best evidence coming from a pre-registered 172-request test.
• Discrepancy noted: reports describe GDPS variously as “ten real-scene tracks” released at the opening and “9 tracks / 23 events” in day-two coverage; both are cited rather than reconciled. AgiBot’s market share is IDC data relayed through Chinese media, not independently published IDC analysis.
• Dedup: The AI HOT window surfaced only the OpenAI cache-side-channel story and Kimi K3.1 Pro, both covered yesterday; the AAIF/Agent Skills cluster was covered Oct 22–23 and appears here only as Radar context because today’s Ace/Infinity stories originate from that same event. Unitree’s limit-up, iFlytek 1024, Qoder on LoongArch, Node2Shell, Claude Duet’s safety architecture and OpenAI’s offline training post were all covered Oct 22–24 and are not repeated. Agent Compute Engine (Oct 22) and Infinity (Oct 23) are reported today because their detailed architecture write-ups and second-day analysis landed in this window.
• Terminology note: /goal in item 1 is Antigravity’s slash-command mode for preserving an objective across context compaction, not a shell command.