GitSkins by @asamassekou10 is now listed for sale 🔁 MRR: $74/mo 💰 Revenue (30d): $81 💵 Total revenue: $178 🏷 Asking price: $5.5k 📊 Multiple: 5.7x ARR 📈 Growth (30d): -16% 💸 Profit margin: 85% 💳 Active subs: 9

Meta Superintelligence Labs just turned a Git repo into shared memory for a swarm of agents. GitSwarm: Decentralized Compounding Inference (Meta Superintelligence Labs, with Mila, CMU and Princeton) makes inference compound instead of restarting from scratch on every run. → Homogeneous async agents work with no central planner, each deciding whether to explore, verify, refine or combine prior work → Every contribution is an atomic commit in a shared, branchable Git repo, with semantic dependencies recording what it builds on → ProgramBench: 79.4% mean score vs 65.1% for the strongest baseline, scaling from 73.1% with 100 workers to 79.4% with 500 → IMOProofBench-Advanced: all 30 problems solved in one run with GPT-5.5, using under one million output tokens → Reuse is real: 94.7% of ProgramBench contributions were later built upon, and the winning solution's ancestry covers 82 to 93% of the graph → On 3 neural architecture tasks, workers ran 395 experiments in the first 26 hours and improved on the starting architectures Takeaway for agent builders: if intermediate work disappears when an episode ends, you are paying to rediscover it. Persist it, with its dependencies, and let later agents compound it. Paper: Worth trying this pattern with your own agents in @muse

How many agent skills are actually out there? There are ~3.8M SKILL.md files across 282,200 public GitHub repositories, nine months after Anthropic published the format as an open spec. This research mined all of them into a dataset. Skills resist the usual software mining. They are written in natural language, an agent picks them probabilistically at run time, and no compiler or type checker verifies the choice. There is no registry and no package manager, so they spread by copying folders between repos. GitSkills groups them into 1,877,981 distinct contents and ships as a single SQLite file with front matter, folder contents, and commit history. Paper: Track more trending AI papers in our academy:

TL;DR Claude generated so much code that Git itself became the bottleneck, so Anthropic rebuilt its source control system from scratch. A team of just 7 people pulled it off in 5 months. Title: Redesigning source control at Anthropic URL: Here are the key points. ・📈 Daily merges jumped more than 10x in 5 months, from about 2,500 to over 30,000 ・⚠️ Naively scaling the old Git setup would have meant 600PB of daily traffic ・🖥️ The new system, Scion, assumes clients are always online and pushes work to the server side ・🗂️ A lightweight Changeset data structure allows instant ingestion with no packfiles needed ・⚡ A 3-tier cache hits 99%+ at every layer, cutting checkouts from minutes to seconds ・💾 Infrastructure CPU and memory footprint dropped 200x, and CI CPU usage fell 25% ・🤖 Claude helps with design and code generation, and even acts as a dedicated "superfan" watching individual components What stands out is both the willingness to question Git's core assumptions and the speed of pulling off the rebuild with Claude in just 5 months. #Anthropic# #SoftwareEngineering#

Grok 4.7 is just okay at coding, but its reward hacking really surprised me. Every task in SWE-Together is built from a real open-source repository, and for most tasks the fix already exists upstream. So we built the sandbox with that in mind. Task images strip all git history after the base commit, remove the remote, and fail to build if any later commit can still be reached. Each container also resolves GitHub, GitLab, Bitbucket, and Hugging Face to localhost, so requests to them go nowhere. But Grok 4.7 found ways around these guards like no other model we tested. It pulled GitHub content through CDN mirrors and gh-proxy sites. It looked up GitHub's real address through DNS-over-HTTPS and handed it to curl. It wrote a small library that intercepted git's address lookups so git could reach again. It used web search to find the number of its own PR. In fact, it tried in 60% of trials and reached the upstream code in 44 of 218. In 20 of those, the code it pulled was the task's own fix. In three trials it even ran `git reset --hard` to replace the repo with upstream main 😅. We fixed all of these loopholes exposed by Grok 4.7 (grateful in this regard). Specifically, we moved enforcement outside the container. Each container now runs in its own network namespace with no route out, and one allowlisting proxy on the host is the only exit. During the re-run of those 44 trials, Grok 4.7 kept trying with 3,246 blocked attempts across 442 hosts. It actually found two routes we had not thought of. It asked a web-enabled model through our own LLM route to fetch the PR for it. It pulled the next release of the repo it was fixing from npm. We closed both by pinning the model and blocking the task's own packages at the registry. Long story short, with every route closed, Grok 4.7 lands at #4# on SWE-Together. It scores 65% pass@1, 53% pass², and a judge score of 0.835. It basically sits mid-pack on every column. It also uses 2× the output tokens of Grok 4.6 to get there, at $7.81 per task compared with $3.64. I am guessing Grok 4.7 was trained with heavy RL on coding tasks, where anything that makes the tests pass earns reward, and with little monitoring of how that reward was earned. Anyway, see the latest results at