Git for a thousand agents

By Mark 9 min read 0 views

😁 Hello, super humans! Yesterday physical AI clocked in; today the agent era came for the one tool you actually live in: version control. Cursor used its first-ever conference to unveil a Git platform built for swarms of bots instead of humans, a 3-billion-parameter model started beating flagships at coding, and SK hynix slid the next memory generation into the AI pipeline. Let’s dig in.

📰 Quick Signals

  • 🧠 AI: Weibo’s open-source VibeThinker-3B, a 3B model built on Qwen2.5-Coder, posts flagship-level math and coding scores and runs in about 6GB of VRAM (VentureBeat).
  • 🤖 Robotics: Beijing’s Booster Robotics tested a humanoid on a soccer pitch and its T1 kicked the ball hard enough to dent a wall, reviving safety questions for a proposed RoboCup (Futurism).
  • 💻 Programming: GitHub made Claude Opus 4.8 generally available across Copilot in VS Code, JetBrains, Xcode and the CLI, with usage-based billing now live (GitHub Changelog).
  • ⚡ Electronics: Snap debuted Specs, a $2,195 see-through AR computer, betting that glasses are the next platform after the phone (Snap Inc.).
  • 📡 Telecom: Taiwan’s ITRI signed an MoU with the European 6G-SNS association and the Taiwan 6G Industry Forum to align trials and standards ahead of a 2030 rollout (VoIP Review).

🔍 The Big Story: Git for a thousand agents

For fifteen years the unit of version control was a human who commits a few times a day. This week Cursor bet that assumption is about to break, and built the alternative.

What happened: At its first-ever Compile conference, Cursor unveiled Origin, a Git hosting platform designed from the ground up for AI agents rather than people (BigGo Finance). The forge is built on technology from Cursor’s Graphite acquisition and tuned for the high-frequency, parallel commits that fleets of agents produce. It is waitlist-only now, with a launch planned for the fall (AlphaSignal).

The details: The live demo is the pitch. A single repo absorbed roughly 22.6 commits per second, about 296,000 clones and 81,000 pushes per hour, with global sync under 400ms and automatic failover in under 10ms. The headline feature is an AI-powered merge-conflict resolver that settles collisions the moment hundreds of agents branch off the same files, plus first-class API and MCP extensibility so the agents talk to the forge directly instead of pretending to be a human with a keyboard. GitHub optimizes for a person reading a diff; Origin optimizes for a scheduler fanning work across machines.

Important

Our take: The interesting claim here is not “GitHub killer,” it is that the merge conflict becomes the real bottleneck once agents outnumber engineers on a repo. If your team is already running parallel agents, you have probably felt it: the model is fast, but reconciling five branches that all touched auth.py is still a human’s afternoon. Whoever automates that reconciliation cleanly owns the next decade of dev infrastructure. You do not need to wait for Origin to prep, though; start running agents in isolated worktrees today (see Code Corner) so the parallel workflow is muscle memory before the tooling catches up.

🗞️ More News

A deep AI-policy and silicon day, with a full roundup across the rest of the beat below.

🧠 AI

  • Nous Research’s open-source Hermes agent got a Stripe wallet, so it can buy products on the web, pay per-call APIs, and even spin up and bill its own SaaS, with your real credentials never entering its chat history (Stripe).
  • A US export-control order forced Anthropic to disable Fable 5 and Mythos 5 worldwide within about 90 minutes, and at the G7, leaders floated a “trusted partners” pact giving allies priority access to top US models (Nextgov/FCW).
  • The EU’s cybersecurity agency ENISA will meet Anthropic on June 19 as the export lockout intensifies Europe’s scramble for AI sovereignty (AI News).
  • Washington held off blacklisting DeepSeek, ChangXin Memory, and more than 100 other firms despite security concerns, opting to keep building the chip supply chain at home instead (TS2).
  • Microsoft is weighing a self-hosted DeepSeek model for Copilot to cut enterprise AI costs by as much as 70% as it shifts to usage-based pricing (Seeking Alpha).
  • Jeff Bezos told VivaTech that AI will create “labor scarcity,” not job loss, arguing demand for human work will outrun supply (The Hill).
  • A WordPress VIP survey of 2,000 people found 60% of US consumers say “AI” in brand messaging is a turnoff, even as firms chase AI-search visibility (TechCrunch).
  • OpenAI is preparing GPT-Bidi-1, a bidirectional voice model that listens and speaks at once so ChatGPT can handle interruptions and adjust mid-sentence (TestingCatalog).
  • SandboxAQ won a definitive $500M CHIPS award to develop AI-aided materials discovery for semiconductor manufacturing, with Commerce taking a minority equity stake (TS2).

🤖 Robotics

  • LG Electronics is building a “training-data factory” in Seoul to mass-produce the physical task demonstrations humanoid robots need most (The Robot Report).
  • NVIDIA’s GEAR Lab, with CMU and UC Berkeley, revealed ENPIRE, a framework where coding agents taught robot fleets to install GPUs into motherboards at a 99% success rate, and it is going open source (Tom’s Hardware).
  • Neura Robotics raised a Series C worth up to $1.4B from Tether, Qualcomm, Amazon, Nvidia, Bosch, Schaeffler, and the European Investment Bank (CNBC).
  • Figure’s BotQ factory is now producing the Figure 03 at roughly one robot per hour, while Boston Dynamics’ electric Atlas began initial deployments (KraneShares).
  • Agility Robotics’ Digit now has seven-plus units active at Toyota Canada under a robotics-as-a-service deal, a sign pilots are turning into contracts (Robotics Center).
  • The State of Robotics 2026 report pegs the market at $38B, up 34% year over year, the sector’s fastest growth in a decade (Robotics Center).

💻 Programming

  • VS Code 1.118 shipped bigger Copilot agent workflows, including remote control for CLI sessions, semantic codebase search, stronger enterprise controls, and lower token usage (VS Code).
  • GitHub’s Octoverse data shows TypeScript has overtaken Python as the platform’s most-used language as AI feedback loops reshape how code gets written (GitHub Blog).
  • Bun keeps eating Node’s lunch: now shipping inside Cursor and Midjourney, with benchmarks showing roughly 3x faster cold starts than Node.js (Bun).
  • Deno 2.7 stabilized the Temporal API for sane date/time handling and tightened npm compatibility, sharpening the runtime fight (Deno).
  • At Compile, Cursor, now operating under its SpaceX acquisition, said it is training a 1.5-trillion-parameter model from scratch on more than 100,000 xAI GPUs (BigGo Finance).
  • GitHub’s usage-based billing for Copilot went live for all users on June 1, replacing flat multipliers with metered premium requests (GitHub Changelog).

⚡ Electronics

  • SK hynix shipped 12-layer HBM4E samples to major AI customers today, hitting 16Gbps per pin with over 20% better power efficiency and 17% better heat resistance than HBM4 (SK hynix).
  • TSMC and partners detailed an HBM4/HBM4E architectural shakeup, with 3nm base dies targeting a 2.5x performance boost and up to 12.8GT/s by 2027 (Tom’s Hardware).
  • Intel Foundry chief Naga Chandrasekaran said the company stays “committed to leading-edge process innovation,” though 18A-P still has to prove yields and demand as Intel courts an Apple deal (TS2).
  • Tensordyne expects more than $200M in orders for its Napier AI-inference system, fabbed by TSMC and developed with Broadcom and Juniper (TS2).
  • Apple is developing camera-equipped AirPods for a late-2027 launch, using the cameras to feed Siri “Visual Intelligence” rather than to shoot photos (GSMArena).
  • The memory build-out has a counterweight: SK Hynix plans to double overall production capacity over five years as AI demand drives a global memory shortage (EE Times).

📡 Telecom

  • US regulators are still hunting for 6G spectrum, with the latest moves focused on freeing midband airwaves for next-generation networks (Light Reading).
  • The GSMA warns 6G will need roughly three times more midband spectrum than is available today, around 2-3GHz per country by 2035-2040 (Telecoms Tech News).
  • Mobile operators, via the NGMN, are pleading not to repeat 5G’s mistakes with 6G, asking for a simpler standardization path to avoid market confusion (The Register).
  • Broadcom launched its BroadPeak SoC to power 5G-Advanced and 6G networks, pushing more silicon into the radio access edge (The Fast Mode).
  • The year ahead in airwaves is shaping up around spectrum sales and squabbles, with regulators and carriers jockeying over who gets what (Light Reading).
  • Nokia laid out how spectrum decisions made now will shape the road to 6G, framing band choices as the real determinant of capacity (Nokia).

👨‍💻 Code Corner

Today’s Big Story is about many agents committing in parallel; you can rehearse that workflow locally right now with Git worktrees, which give each agent its own working directory and branch off one shared repo so they never trample each other’s files:

# Give three agents isolated worktrees off the same repo
for agent in fix-auth add-tests refactor-api; do
  git worktree add "../wt-$agent" -b "$agent"
done

git worktree list

# When an agent's branch is ready: merge, then clean up its worktree
# git merge add-tests && git worktree remove ../wt-add-tests

Tip

Worktrees solve the “isolation” half of parallel agents; they do not solve the “merge five branches that all touched the same file” half. That reconciliation pain at scale is exactly what Cursor’s Origin is betting it can automate, so feel where it hurts now and you’ll know what to evaluate later.

🧰 Toolbox

  • VibeThinker-3B is an MIT-licensed 3B reasoning and coding model that fits in about 6GB of VRAM, or 2GB when quantized.
  • Ollama lets you pull and run open models like VibeThinker locally with a single command and no API key.
  • ElevenLabs turns a script into an expressive voiceover in 70-plus languages, with voice cloning and editor export.
  • Bun is a fast all-in-one JavaScript runtime, bundler, and test runner now shipping inside Cursor and Midjourney.
  • Stripe agent toolkit gives an AI agent its own wallet to buy products and pay per-call APIs, with human approval gates.

🔌 Component of the Week (rotating)

NVIDIA Jetson Orin Nano Super — A pocketable AI dev kit that has become the obvious home for the wave of tiny-but-capable models like today’s VibeThinker-3B. The Super refresh pushes the module to around 67 INT8 TOPS with 8GB of LPDDR5, enough to run small reasoning models and vision pipelines fully on-device with no cloud round-trip, which is why it keeps showing up in robotics and edge-AI builds. It runs the full CUDA and JetPack stack, so the same code you prototype on a workstation usually drops straight onto the board. At about $249 for the dev kit it is the cheapest serious on-ramp to local physical AI. Grab the developer kit details to see the I/O and power modes.

😀 The Bot Says…

A repo took 22 commits a second today; robots taught themselves to slot GPUs, and Apple wants to put cameras in your ears so Siri can finally see your fridge. The machines are shipping faster than we can review the diffs, but at least one of them still kicked a soccer ball straight through a wall. 🤖⚽

That’s all for today! If the merge conflict is the new bottleneck, what part of your workflow would you hand to an agent first? Reply and tell us.