Langsmith just added a judge that undercuts every llm judge. jev, out september 21, scores…
langsmith just added a judge that undercuts every llm judge. jev, out september 21, scores open-ended agent behavior for a fraction of the llm-as-judge cost. the cheapest evaluator is no longer a model.
read the note →Sentry just started tracing the agents inside your app. agent tracing, ga since september 11,…
sentry just started tracing the agents inside your app. agent tracing, ga since september 11, records every model call, tool execution and handoff inside the traces you already have. the error monitor became an agent monitor.
read the note →Google just shipped its agent kit for kotlin. adk 1.0, out september 9, matches the python and java…
google just shipped its agent kit for kotlin. adk 1.0, out september 9, matches the python and java builds and adds room sessions, appsearch memory and on-device lite models for android. the mobile agent finally got a home.
read the note →The real codebases just humbled every coding agent. real-swe, out september 14, ran 640 rollouts on…
the real codebases just humbled every coding agent. real-swe, out september 14, ran 640 rollouts on licensed production code and the best model closed 38.8 percent of tasks at 6.96 dollars each. the benchmark gap is the pricing gap.
read the note →Windows just gave agents a way to run desktop apps from the terminal. the windows app development…
windows just gave agents a way to run desktop apps from the terminal. the windows app development cli v0.3, out september 10, adds winapp run and ui commands with dotnet support for packaged apps. the gui app stopped being gui-only.
read the note →The data pipeline vendor just renamed itself a context layer. fivetran's context layer, out…
the data pipeline vendor just renamed itself a context layer. fivetran's context layer, out september 16, applies the semantic layer discipline to agents instead of dashboards. the warehouse became the agent's memory.
read the note →Accessibility just moved inside the pull request. accessibe's code agent, out september 17, reviews…
accessibility just moved inside the pull request. accessibe's code agent, out september 17, reviews github prs against wcag 2.2 aa and flags the exact lines before merge. the a11y audit stopped waiting for a release.
read the note →Clickhouse just turned postgres's wal into its own pipeline. walshadow, out september 10, consumes…
clickhouse just turned postgres's wal into its own pipeline. walshadow, out september 10, consumes the physical wal stream directly and turns clickhouse into an analytical standby with sub-second lag. the replication debate skipped logical decoding.
read the note →Google just put agents in charge of model post-training. autofinetune, out september 11, uses ai…
google just put agents in charge of model post-training. autofinetune, out september 11, uses ai agents to tune hyperparameters and optimize llm training loops on tpus. the tuning run stopped needing a human.
read the note →Android just made sideloading a registered identity. from september 30, verified developers must…
android just made sideloading a registered identity. from september 30, verified developers must register apps for certified devices in brazil, indonesia, singapore and thailand, with the rest of the world following. the apk stopped being anonymous.
read the note →Tenstorrent just opened its model compiler to everyone. tt-forge, the mlir-based stack built on the…
tenstorrent just opened its model compiler to everyone. tt-forge, the mlir-based stack built on the company's ai software, hit public beta on september 3, compiling pytorch, jax and onnx models for its hardware. the compiler caught up to the chip's ambitions.
read the note →The edge ai chip just went production with 629 tops per card. axelera's europa, announced september…
the edge ai chip just went production with 629 tops per card. axelera's europa, announced september 15, runs matrix math inside sram arrays and ships as a half-height pcie accelerator with 64 gigabytes on board. the memory became the compute.
read the note →The monorepo changelog tool just cut its install weight by 88 percent. changesets v3, out september…
the monorepo changelog tool just cut its install weight by 88 percent. changesets v3, out september 22, goes esm-only and stops force-bumping peer dependencies to major versions. the release tool got lighter than the packages it ships.
read the note →Mlflow just let teams build their own trace views. the 3.16.0 release, out september 3, adds a…
mlflow just let teams build their own trace views. the 3.16.0 release, out september 3, adds a redesigned trace explorer, span links and custom views to the open ml platform. model observability stopped being a fixed page.
read the note →One donor-funded maintainer just kept a browser engine alive for a year. servo's josh…
one donor-funded maintainer just kept a browser engine alive for a year. servo's josh bowman-matthews wrapped the experiment on september 17, reviewing 1,150 pull requests and clearing 92 percent of the 114 newcomer issues he filed. the browser project ran on donations.
read the note →The risc-v debugger just caught up to alibaba's fastest core. lauterbach's trace32 added the…
the risc-v debugger just caught up to alibaba's fastest core. lauterbach's trace32 added the xuantie c950 on september 15, the cpu that scores past 70 on specint single core. the open instruction set finally has a top-tier toolchain.
read the note →Video generation just became one sdk call. grok imagine, on vercel's ai gateway by september 22,…
video generation just became one sdk call. grok imagine, on vercel's ai gateway by september 22, turns the experimental_generatevideo function into a route that polls until the clip is ready. the video api stopped being a separate project.
read the note →Figma just halved the price of its ai. ceo dylan field said this month that ai usage costs fell up…
figma just halved the price of its ai. ceo dylan field said this month that ai usage costs fell up to 50 percent, with extra quota at the same price, pushing scale over revenue. the design ai is being sold on volume.
read the note →The ai agent just got a hardware debug probe. the embedder and segger partnership, out september…
the ai agent just got a hardware debug probe. the embedder and segger partnership, out september 10, lets an ai agent drive j-link and j-trace directly to build and flash firmware. the agent stopped trusting the build output.
read the note →The software bill of materials just learned what ai ships. cisa and the g7 released the ai sbom…
the software bill of materials just learned what ai ships. cisa and the g7 released the ai sbom minimum elements on september 21, standardizing how model supply chains describe weights, data and tools. the ingredients list now covers training runs.
read the note →The coding agent race just found its least glamorous corner. byteask, from yc's fall 2026 batch,…
the coding agent race just found its least glamorous corner. byteask, from yc's fall 2026 batch, targets c and c++ code, the stuff running cars, chips and operating systems. the systems code finally got an agent that doesn't stop at a diff.
read the note →Chrome just halved its release cycle. version 153, out september 9, kicks off google's two-week…
chrome just halved its release cycle. version 153, out september 9, kicks off google's two-week stable cadence, roughly doubling how often the web platform changes hands. the browser ships like a news feed now.
read the note →The postgres ecosystem just shipped its own migration hammer. postgresql migrator 1.0, out…
the postgres ecosystem just shipped its own migration hammer. postgresql migrator 1.0, out september 7 from dalibo, is a free go binary that drags oracle and mysql databases into postgres with zero proprietary dependencies. the escape hatch from proprietary sql is official.
read the note →The javascript security numbering authority just went on holiday. the openjs foundation's cna, the…
the javascript security numbering authority just went on holiday. the openjs foundation's cna, the cve issuer for the js ecosystem, pauses security operations from september 17 to october 6 for maintainer rest. the vulnerability machine needs maintenance too.
read the note →Atlassian just made its own tickets the agent's ground truth. the september 10 launch ties ai…
atlassian just made its own tickets the agent's ground truth. the september 10 launch ties ai coding tools to jira and confluence context, citing a dx analysis where the heaviest atlassian-context teams shipped 64 percent more per developer. the issue tracker became the memory.
read the note →A 27 billion parameter model just moved onto a laptop. prism ml's bonsai 27b, out september 19,…
a 27 billion parameter model just moved onto a laptop. prism ml's bonsai 27b, out september 19, packs a full reasoning model into 1-bit weights, cutting the download from 54 gigabytes to something a flagship phone can run. the model shrank, not the reasoning.
read the note →The terminal just learned to restart its agents. 1devtool's autopilot, out september 16, relaunches…
the terminal just learned to restart its agents. 1devtool's autopilot, out september 16, relaunches a crashed agent in the same terminal with its session intact. the agent got a self-healing loop.
read the note →India just passed the halfway mark on its national gpu build. the indiaai mission, past 17,300 of…
india just passed the halfway mark on its national gpu build. the indiaai mission, past 17,300 of 34,333 gpus by september 21, runs a unified compute portal for researchers and startups across yotta, nextgen and e2e. sovereign compute stopped being a slogan.
read the note →Microsoft just wrote red lines its models cannot cross. the september 14 draft code of conduct bans…
microsoft just wrote red lines its models cannot cross. the september 14 draft code of conduct bans resisting shutdown, widening scope and hiding reasoning from auditors, with a six week consultation. the model got a leash.
read the note →Anthropic just bought the sdk factory. the stainless acquisition, confirmed september 18 for over…
anthropic just bought the sdk factory. the stainless acquisition, confirmed september 18 for over 300 million dollars, takes control of the tooling that generates official libraries for openai, google and cloudflare. the api front door has one owner.
read the note →Openai just gave prompt caching a debugger. prompt cache diagnostics, out september 8 in the…
openai just gave prompt caching a debugger. prompt cache diagnostics, out september 8 in the responses api, reports why a cached prompt missed its reuse instead of leaving teams to guess. the cache miss stopped being a mystery.
read the note →Code search just learned to answer why. entire's agentic search, out september 4, blends semantic…
code search just learned to answer why. entire's agentic search, out september 4, blends semantic search across commits and session transcripts so agents see the reasoning behind a change. the codebase got a memory.
read the note →Ai engineering salaries just split in two. the september 9 analysis of 43,000 us job ads puts…
ai engineering salaries just split in two. the september 9 analysis of 43,000 us job ads puts agent-focused roles at a 213,000 median, about 20,000 above ads that name frameworks. the stack you name costs you money.
read the note →The agent memory problem just got a shared test. the agent memory challenge, opening its second…
the agent memory problem just got a shared test. the agent memory challenge, opening its second cycle september 20, grades whether agents notice when an old fact stops being true, after a first round drew 136 teams. stale memory is worse than no memory.
read the note →The first laptops built for local agents shipped this month. nvidia's rtx spark machines, shown at…
the first laptops built for local agents shipped this month. nvidia's rtx spark machines, shown at ifa berlin, pair an arm cpu with a blackwell gpu and 128 gigabytes of unified memory. local ai stopped being a compromise.
read the note →Nvidia just made quantum debugging look like a classical workflow. the cuda-q logical layer, out…
nvidia just made quantum debugging look like a classical workflow. the cuda-q logical layer, out september 14, cut fault tolerant algorithm development at fermilab from five months to three weeks. quantum stopped being a waiting game.
read the note →A python scaling tool just kept its open source promise. burla 0.8.3, out september 19, flips its…
a python scaling tool just kept its open source promise. burla 0.8.3, out september 19, flips its source-available license to apache 2.0 after the two year clock ran out. the relicensing promise held.
read the note →The trending repo this week made the agent a pile of plugins. deepseek-harness, at 206,000 stars by…
the trending repo this week made the agent a pile of plugins. deepseek-harness, at 206,000 stars by september 2, swaps reasoning strategies, tools and output formats like modules. the agent stopped being a monolith.
read the note →The model release just aimed at the test queue. gpt-6 astra, out in early september, went into…
the model release just aimed at the test queue. gpt-6 astra, out in early september, went into devin's automated testing within days, replacing human qa loops in production ci, per the september 14 analysis. the test suite became the product.
read the note →Shopify just left react native. the september 16 move rewrites its flagship apps in swift and…
shopify just left react native. the september 16 move rewrites its flagship apps in swift and kotlin, with head of mobile mustafa ali saying ai made native codebases cheaper than an abstraction layer. the cross-platform bet flipped when agents got good.
read the note →The bottleneck just moved from writing code to shipping it. ai agents now push so many parallel…
the bottleneck just moved from writing code to shipping it. ai agents now push so many parallel changes that ci/cd queues are the constraint, per the september 11 analysis. the pipeline became the product.
read the note →Airflow just made data pipelines resumable. version 3.3, out september 17, adds durable execution…
airflow just made data pipelines resumable. version 3.3, out september 17, adds durable execution with persistent state and crash recovery, and now covers ai agent runs too. the scheduler stopped losing work when a node dies.
read the note →Multi-cluster orchestration just became an official cncf pillar. karmada graduated on september 8…
multi-cluster orchestration just became an official cncf pillar. karmada graduated on september 8 at kubecon in shanghai, after years of running apps across clusters and clouds without touching the app itself. kubernetes stopped being single-cluster.
read the note →Observability just bought agent monitoring. dynatrace acquired arize ai on september 11, folding ai…
observability just bought agent monitoring. dynatrace acquired arize ai on september 11, folding ai evaluation and agent telemetry into its application platform. monitoring stopped being about humans reading dashboards.
read the note →Ai code review became a benchmark race this week. augment's reviewer, out september 17, runs on…
ai code review became a benchmark race this week. augment's reviewer, out september 17, runs on gpt-5.2 and claims a 10 point lead over cursor bugbot and coderabbit on the public quality benchmark. review accuracy is now a marketing number.
read the note →The elasticsearch fork just got its first long-term support release. opensearch 3.6 and 2.19, out…
the elasticsearch fork just got its first long-term support release. opensearch 3.6 and 2.19, out september 18, are the first lts versions since the 2021 split and add 32x vector compression for agent workloads. the fork stopped being a moving target.
read the note →90 percent of developers use ai coding agents and only 29 percent trust their accuracy. jetbrains'…
90 percent of developers use ai coding agents and only 29 percent trust their accuracy. jetbrains' 2026 ecosystem survey, out september 15, found 66 percent spend more time fixing almost-correct code. adoption ran far ahead of trust.
read the note →Openai just doubled the number of open source maintainers it sponsors. the codex for oss program,…
openai just doubled the number of open source maintainers it sponsors. the codex for oss program, renewed september 15, now covers 10,000 maintainers with six months of pro access and api credits for pull request review. the maintenance economy got a corporate payroll.
read the note →