AI Daily Digest · Oct 26, 2026 Evening (Issue #2)

> Focus: AI Coding × Embodied Intelligence. Cross-checked against the Oct 23, Oct 25, and this morning's issue #1 — no duplicate content. Only related item: the PHP2Shell vulnerability was previously covered from the discovery side; this issue covers only new remediation-side developments.

---

I. Top Stories (5)

1. [AI Coding / Infrastructure] Git is breaking under agent load: Anthropic replaced its own Git with 7 people in 5 months

The event: Anthropic published an engineering blog post, "Redesigning source control at Anthropic," detailing why it built a custom version control system, Scion, and migrated its entire monorepo.

Hard numbers (all from the official blog):

• Daily merged changes rose from ~2,500 in May to 30,000+ in September (12x in 4 months), with Claude writing over 80% of the code;
• Main branch history is near 100 GB, compressed working copies exceed 25 GB; CI alone performs ~25 million code checkouts daily — naive Git pulls would equal 600 PB of Git traffic per day;
• Solution: a 7-person team completed the Scion migration in 5 months, and also built the Antique code review tool, code search, and an access control system;
• Results: infrastructure footprint (CPU+memory) is 200x smaller than the original Git mirror cluster, checkouts are two orders of magnitude faster, and CI reclaimed 25% of CPU capacity.

The most counterintuitive finding: Scion concluded that "merge queues make no sense at agent scale" — when hundreds or thousands of changes can merge while tests are still running, the traditional "queue, verify, then merge" model breaks down, forcing a shift to test-after-merge, before-deploy. Git's advantage — designed for distributed offline work — becomes pure cost when clients are almost entirely always-on agents and CI jobs.

Why it matters: 1. This is the hardest empirical evidence yet of "agent-scale code volume": when AI writes over 80% of code, the first thing to break isn't the model or the IDE — it's developer infrastructure designed for human cadence (version control, CI, review); 2. Anthropic explicitly says Scion won't be open-sourced and isn't suited to general use cases, but recommends GitHub/GitLab learn from its approach — companies heavily using coding agents will hit the same wall within 6-12 months; 3. Direct lesson for platform/infra teams: change velocity itself becomes a new system constraint, and review/verification architecture (up-front gating vs. post-merge regression) needs rethinking in advance.

2. [AI Coding / Runtime] AWS open-sources Infinity: agents spend 90% of their lives "waiting" — why pay by the hour?

The event: On Oct 23, AWS together with UC Berkeley (the Hydro project) open-sourced the Infinity runtime (Apache 2.0, Rust) — bringing the "suspended coroutine" concept from programming languages to agents: when an agent issues a tool call, it suspends and retains only conversation state, resuming when results arrive.

Hard numbers:

• Each idle agent occupies only ~100 KB of memory; official figures say 8 GB can host ~75,000 concurrent agents (comparable to a Raspberry Pi 5's memory; note: tests used a simulated model + single-threaded EPYC, not independently verified);
• Naturally serverless-friendly: during waits, Lambda functions can simply exit — a tool call can wait three days at zero compute cost;
• Clever technical details: "synthetic tool calls" (receive_event__injected) push async events back into the conversation without modifying existing model APIs; inter-agent communication uses spawn_thread / report_to_parent to avoid child tasks flooding the parent context; MCP servers run unchanged through a compatibility layer.

Why it matters: 1. Agent cost structure is explicitly splitting into "compute" and "wait" — in reality, coding agents may spend ~90% of their lifecycle waiting on builds, reviews, and approvals. Billing a resident process means paying for queue time. Infinity changes the billing unit from "hours" to "events"; 2. Google's Agent Compute Engine, released the same week, takes the other path: microVM hard isolation with credentials kept out of the sandbox. One makes "waiting cheap," the other makes "running safe" — real production stacks will likely stack both layers; 3. Teams building multi-agent platforms can directly borrow the "slicing + message queue + idempotent tools" architecture — race conditions haven't disappeared, just moved to per-thread message-ordering guarantees.

3. [AI Coding / Security] Cognition open-sources a "vulnerability-fix skill pack": one SKILL.md, working across 70+ coding agents

The event: On Oct 25, Cognition AI open-sourced devin-vuln-fixes (MIT license) — just three SKILL.md files that enable Devin, Claude Code, Codex, Cursor, Copilot, Gemini CLI, and 70+ other coding agents to automatically fix this week's most severe vulnerabilities: self-hosted GitLab CVE-2026-110264 (CVSS 10.0, the maximum score), PHP2Shell (unauthorized RCE across PHP web SAPIs — the remediation follow-up to a previously covered vulnerability), and Node.js Brotli RCE.

The workflow design is worth more attention than the vulnerability list (five steps hardcoded into the skill files): 1. Read-only inventory of Dockerfiles / CI configs / IaC, outputting an affected-assets table; 2. Stop at a human approval gate — no file changes without approval; 3. All fix versions are pinned (e.g., PHP 8.5.12, GitLab 19.5.1), and artifacts containing fixes are confirmed to actually exist before acting — "guessing version numbers" is explicitly forbidden, eliminating hallucinated patches; 4. Run before/after tests on production images; new alerts count as regressions; 5. One focused PR per repository, with a rollback plan.

Why it matters: 1. The coding agent battleground is shifting from "writing new code" to "fixing production incidents" — and the latter lacks not generation capability but trust. Encoding approval gates, real verification, and no-guessing-versions into skill files essentially demotes AI from autopilot to "a junior engineer who can sign off but must be reviewed"; 2. The distribution method carries a bigger signal: this is the first time the Agent Skills standard (donated by Anthropic to the Agentic AI Foundation under the Linux Foundation, alongside MCP) has been used to deploy capabilities across vendors — the same skill runs in five competing tools, bypassing the plugin ecosystem's moat; 3. One caveat: no official fix-rate benchmark has been published; managed runtimes/devices/Electron apps are only flagged, not fixed — the tool's value ceiling depends on your own test suite.

4. [Embodied AI / Models] GPT-6 Astra drives a robot arm with no fine-tuning: general-purpose LLMs start crossing into the physical world

The event: Citing multiple tests, a Huaxi Securities Oct 26 weekly report states that GPT-6 Astra, without fine-tuning, controls a robot arm purely through reasoning — topping a zero-shot robot manipulation benchmark at 69 points, with 95% real-robot grasp success.

The other side must be seen clearly: high-precision task success rate is only 10%, with markedly longer completion times. The replacement of the "specialized cerebellum" by a "general brain" has so far only covered pick-and-place tasks; precision assembly remains far off.

The concurrent tech stack is converging (from the same weekly report):

• NVIDIA's GTC made Physical AI the main theme, proposing "compute is data": releasing a data factory blueprint, a 14-day-deployable DSX POD infrastructure, the GR00T N1.7 humanoid model, and the Agile ONE SN platform;
• Underlying algorithms are moving toward architectural fusion: NVIDIA × Harvard's Hydra-0 reduces motion error by 90.4% via a unified action-stream interface; DeepRobotics' DR-WM-1.0 introduces physical-property priors for the first time; Agibot's GE-Act 2.0 integrates prediction and generation, raising zero-shot success in unfamiliar scenarios to 44.1% with 30,000 hours of data.

Why it matters: 1. If "general models controlling real robots zero-shot" is independently replicated, it will shake the landscape of "every robotics vendor building its own VLA small model" — the specialized-model moat is being breached head-on: the base model's general capability is eating the premium on peripheral customization; 2. But the 95% vs. 10% split shows: grasping is the passing line, precision manipulation is the moat. Specialized models won't die soon — they'll retreat to the high-precision zone; 3. "Compute is data" means the competitive focus shifts to real-robot data collection capacity — data collection itself is becoming a business.

5. [Embodied AI / Industry] Toyota officially enters, safety certification fills in: humanoid robotics shifts from "startup civil war" to "giants' all-out war"

The event: Three signals converged this week —

• Toyota formally joins the humanoid robot race (Financial Times, Oct 26): automotive manufacturing, supply chain, and motion-control expertise is accelerating its migration into robotics;
• TÜV SÜD × Accenture × NVIDIA jointly advance safety certification for AI robots in factories/warehouses (Oct 25): as autonomous decision-making enters industrial robot systems, "verifying robot behavior" is becoming a distinct market segment;
• UBS mapped 87 global humanoid robotics industry players: from NVIDIA to CATL, competition has expanded from complete machines to chips, batteries, actuators, models, sensing, and manufacturing supply chains.

Why it matters: 1. Automakers are "natural robot companies" — motors, motor control, structures, supply chains, and cost engineering are all in place. Toyota's entry weighs more than yet another startup funding round; giants are dragging the industry from a "demo competition" into a "manufacturing war"; 2. Safety certification systems are filling in, echoing the governance thread on the AI side (EU AI Act enforcement, FTC forensics): "auditability" in embodied intelligence is no longer a paper concept but a prerequisite for winning orders; 3. The 87-player map itself footnotes the bubble debate: when there are so many participants you need a chart to track them, the next thing to watch is who wins repeat purchases (not first-unit demos).

---

II. Quick Hits

| Event | Key points | Why it matters | |---|---|---| | Qiyuan ships 5,000 personal robots (Oct 26) | Covering companionship, housework, education; positioned as a milestone in "Physical AI consumerization"; a concurrent CAICT report: real success rate on household tasks is only 12% | "Can build 5,000 units" and "can do housework" are two different things — consumerization is running ahead of capability maturity | | MiniMax open-sources M3.1 Flash | 428B total params / 23B active; 100+ trillion tokens processed by real users within 72 hours of launch; Ascend, NVIDIA, AMD integrated day one | Following Kimi K3.1 Pro, the open-source coding/agent track keeps intensifying; the inference economics of sparse attention + W4A4 quantized routing deserve study | | Unitree new robot stumbles on stage, sparking debate (Oct 25) | Robot lost stability during a stage demonstration; social media reactions split | Stage mobility ≠ deployment reliability — the demo arms race bites back | | Anthropic Q3 revenue up nearly 50% to $16.8B | Disclosed pre-IPO; competitor distillation contributed 5-7% of revenue; 80%+ of merged code written by Claude | "The model seller is its own biggest customer" — the flywheel data of AI companies using AI to build AI laid out this completely for the first time | | TUM × NEURA establish robotics fusion center and RoboGym | Shared testing/training facilities connecting research, talent cultivation, and commercialization | Europe is catching up on "training infrastructure," converging with China's "train-compete-produce-fund" loop: the bottleneck is landing on data and training supply |

---

III. One-Line Takeaway

Today's theme is "the bottleneck moving down the stack": model capability is no longer the bottleneck, and constraints are shifting layer by layer. On the coding side, Anthropic proved by replacing Git that change velocity breaks before models do, AWS made agent "waiting" free, and Cognition encoded "trust" into skill files. On the physical side, GPT-6 Astra's zero-shot robot arm control breached the specialized-model moat, and Toyota entered with a century of manufacturing heritage. Every bottleneck solved immediately spawns the next layer's business — infrastructure, runtimes, safety certification, real-robot data. That is AI's true procurement list.