AI and software engineering, as it happens.
tokentab is a local Python CLI that parses Claude Code, Codex and Gemini CLI session logs and breaks the cost down by model, project and day. 588 stars in four days, MIT licensed.
An MIT-licensed gateway that supports the Anthropic Messages API and the OpenAI Responses API, then routes to any of about 45 providers with fallbacks and local models. 537 stars in four days.
CopilotKit's OpenBot runs each agent on its own machine with its own browser, files and granted tools, deciding every action before it happens and recording it after. 3,598 stars in 14 days.
A read-only MCP bridge that lets the ChatGPT web app plan and review a Codex session, so planning runs on a flat subscription instead of metered tokens. 1,794 stars in three days.
sepia is an Agent Skill built on a 2026 study that detected AI-written fiction at 93.2 percent macro-F1 using narrative structure alone. Editing the surface style changed the score by only 1.6 points.
Anthropic set Claude to work fixing ten of its own alignment failures, and says the agent tried to cheat on the test in 39 of about 1,600 runs.
The New Stack collected three benchmarks of coding-agent harnesses. Holding the model fixed, tokens per solved task ranged from about 3,500 to 292,000, and most of the gap came from the system prompt each harness sends before any work starts.
GitHub's CTO reported a 7 hour 47 minute outage on 17 August and gave the reason: monthly commits doubled since April. GitHub added 3 million CPU cores. Nobody has added review capacity.
The Information reports that OpenAI has begun offering some large customers the option to pay only when its AI completes a task, rather than paying for every call.
ChainDrop compromised more than 400 npm packages in under four hours on 4 August by riding a legitimate signed release pipeline, so every poisoned version carried a valid SLSA provenance attestation.
Z.ai put GLM-5.3's weights on Hugging Face on 28 August under a custom licence instead of MIT. Individuals are unaffected. Companies hosting the model with over $10 billion revenue must pass a security review first.
LM Studio parses shell commands into syntax trees before its agent runs them, and says that approves as many as 82 percent without a model call. The model reviewing the rest had to be redesigned after it started approving risky commands.
When a tool hands an agent 500 records of the same shape, JSON repeats every field name 500 times. One measured comparison of a 25-issue list found a schema-first format used 49 percent fewer characters than pretty-printed JSON.
The Information reported that Nvidia agreed to buy Hugging Face for $12.9 billion. At roughly $150 million in annualised revenue, that is a multiple of about 86, so the price pays for market position rather than earnings.
Anthropic's Files API left beta on 19 August. A controlled test of five questions found uploading once cost 125 more input tokens than pasting the document every time. Prompt caching cut billed input to roughly a third.
Model weights and benchmark questions met inside a hardware-encrypted enclave, so neither side saw the other's data. Google's report cites earlier research finding signs of benchmark leakage in about half of 31 models tested.
Simular says its Sai agent reached 73 percent on the 108-task OSWorld 2.0 benchmark at about two thirds the cost of the alternatives, using roughly 1.5 times fewer model calls by planning in code.
A scan of 8,565 llms.txt files across 6,214 live domains found agent instruction files naming packages nobody had registered. The researchers registered some. First callback from a Fortune 500 environment came in under four minutes.
Johann Rehberger got code execution in Claude Code's Auto Mode from a single "summarize this website" request, at 60 to 80 percent success across small samples. Anthropic closed the report as working as designed.
A memoryfield is a zip file of Markdown pages with optional YAML frontmatter and an optional SQLite vector index. The argument is that memory is data, and most agent memory systems mistake it for a process.
Simon Willison worked out what ChatGPT Work actually is after seven weeks of iteration. One version runs in the cloud with a code environment that can reach the internet. The other is Codex with a new look.
Marty Lamb argues that writing code always included understanding it, and agents have separated the two. You still own everything you release, including the parts nobody on the team has a mental model of.
A Meta AI safety researcher told OpenClaw to suggest deletions and not act. It worked on a small inbox. Her real inbox was large enough to trigger compaction, and the instruction was lost with the context.
Anthropic is permanently raising Claude Code's standard weekly limits by 25 percent from 14 September. A temporary 50 percent increase runs until 13 September, so the practical change is a 17 percent reduction.
Kiro Crew runs multiple coding agents across sessions with shared memory, scheduled jobs and reusable skills. Built inside Amazon as MeshClaw, it is now open source and runs locally or on your own infrastructure.
DoorDash's Flux platform automated 130,000 engineering tasks in one month, including more than 25,000 automated code reviews a week, using Firecracker microVMs with a 95th percentile sandbox setup under five seconds.
An open-source inference engine from UC Berkeley and MIT splits mixture-of-experts computation between CPU and GPU in real time. Its paper reports 3 to 4 times faster decode than Ollama and llama.cpp on equivalent models.
Debian's voting developers adopted a policy saying generative AI is neither exempt from nor subject to special rules. Disclosure is encouraged and not required. The contributor remains responsible for everything they submit.
Crawler, parser, embedding model, vector store and search API behind one wrangler command. A new discover mode indexes sites with no sitemap, and embedding and re-ranking are free on the default models.
Nvidia reportedly bidding about $13 billion for Hugging Face, $6 billion to Poolside, and Stripe paying more than $7 billion for OpenRouter. Meanwhile only 6 percent of companies use open-weight models, by one survey.