Coding agents learn libraries from training data with a cutoff date. When a project pins newer versions, the agent keeps writing calls to APIs that were renamed or removed since then. The type errors only show up after the code is written. stalemap is an open-source tool from Umair on GitHub that finds the dependencies in a JavaScript or TypeScript project that changed after a model's training cutoff and tells the agent what changed. It writes a short, ranked list of those breaking changes into the agent's instruction file, can run the project's own type-check after each edit to catch code that hits a changed API, and offers an MCP tool the agent can ask in plain words. The notes come from type diffs, release notes and upgrade guides for the exact versions the project pins, so the agent sees the new API before it writes the old one.

Getting a circuit board ready for the fab is a long chain of manual steps. You check the board summary, wire up the schematic, run the electrical and design rules checks, fix what they flag, and export the Gerber files and the bill of materials. A coding agent can help with the firmware, but all of that board work happens by hand inside KiCad. KiCad MCP Pro is an open-source MCP server from Osman Aslan on GitHub that gives an agent tools for that whole flow. It can read the board summary and the schematic, add wires and labels, run ERC and DRC through KiCad's own engines, and export Gerbers, drill files and a BOM. The project's own first-board guide follows the same order: pick a project, inspect the board, set the outline, build the schematic, run the checks, and export the manufacturing files. The project describes it as a first-pass design and review assistant, and says generated boards and fab files need review by a qualified person before anything gets made. For anyone who already works with an agent on firmware, the board checks and fab exports can now happen in that same session.

Linear's Loops can now start on their own from events in Slack and GitHub. Per the changelog, a loop can answer product questions as they come into a Slack channel, or kick off when a pull request gets a "ready for testing" label. Teams can also save a working loop as a template from its menu, and it shows up next to Linear's built-in templates for everyone in the workspace. Each loop gets its own reasoning level too, with Auto following the workspace default or Low, Medium and High picked by hand. Loops are Linear's automations that respond to changes in Linear and connected tools. Work that used to wait for someone to spot a Slack thread or a PR label now starts the moment the event happens. linear.app/changelog/2026-10…

A recent Grok Build release stops worktrees from asking you the same permission questions all over again. Per the release notes, worktrees now reuse the saved "always allow" answers from the main checkout, so a tool you already approved there runs without a fresh prompt. Any "never allow" answers from the main checkout still apply either way, and setting worktree_inherits_main_approvals to false under [features] turns the inheritance off. The same release makes prompt suggestions respect vim mode, so Tab moves focus to the scrollback when vim mode is on. Grok Build is xAI's coding agent for the terminal. For anyone running a worktree per parallel task, the approvals already given in the main checkout now carry over to every one of them. x.ai/build/changelog

Prompt-to-app tools have mostly produced good-looking front ends with nothing behind them. The moment a site needs to store sign-ups or let someone edit content, it turns into a separate backend project. Figma Make now sets that part up from the same prompt. With Payload, the open-source content framework whose team joined Figma, prompting an app also creates a database, file storage, user login and an admin panel, and Figma hosts it. Figma says the integration is now available to everyone. Per Figma, a designer can prompt the site, a marketer can update its content in the admin panel, and a developer can extend it with Payload's framework, with permissions controlling who can change what. Hosting for these projects is in beta and free for now, and Figma says it will announce pricing before that changes. Because the backend is open-source Payload, a developer can take the app further later without a rebuild from scratch.

Most teams running coding agents end up building their own setup out of compute, harnesses and review flows they already have. Then one hosted platform sits in the middle of all of it, and every change to that piece waits on someone else's roadmap. HumanLayer is a platform for running coding-agent work across a team, with tasks, plans, sessions, diffs and PRs in one place. Its core is now open source under an Apache license, and per the announcement it is the same code the hosted version runs: the desktop and web apps, the local daemon that runs agent sessions, the cloud API, and the GitHub, Linear, Jira and Slack integrations. Builds you run yourself send no telemetry, and for now a few third-party API keys are required to start, which the team says it will make optional first. Once the agents can read and patch the platform that runs them, a team can fix and extend its own setup directly.

Mistral put out another Vibe release, and the big change is how many subagents it will run for you at once. Per the release notes, the default limits go up to thirty-two subagents alive and thirty-two running at the same time, from eight and four before. Two new settings, max_subagents and max_concurrent_subagents, let you set those caps yourself. Subagents whose last turn failed now stay in the list marked as failed, and the model's reasoning shows by default as a collapsible Thought row. Mistral Vibe is Mistral's open-source coding agent, with a command-line tool and a desktop app. Higher defaults mean a big fan-out job, like reviewing a whole repo folder by folder, runs in parallel without touching the config first. pypi.org/project/mistral-vib…

Most voice assistants turn your speech into text, answer in text, and read the reply back. Tone, pauses and laughter get lost along the way. Dessa is a full-duplex speech-to-speech model from Block (@blocks), trained from scratch on recorded human conversation alone. Audio goes in, audio comes out, and it can listen and talk at the same time. Block says the behavior came from listening: it laughs mid-sentence when something is funny and gets louder when asked to speak up, and every demo clip on its page is a live, unsteered recording. For now Block is opening it to a small group of AI safety researchers. That makes it a rare test of what a model can learn from sound alone.

Most video models hand you a finished clip and stop there. Steering a scene as it unfolds usually takes a game engine. Hy-Nomas, an interactive video world model from Tencent Hunyuan (@TencentHunyuan), lets you drive a generated world with text prompts or with camera and keyboard moves, switching between them whenever you want. It keeps a memory of the world, so places and characters hold together over long runs. Tencent released the model weights, inference code, a technical report and an online demo under an Apache license, with inference tuned for consumer GPUs. hy-nomas.hy.tencent.com/

Coding agents write plenty of UI they never click. The build goes green, the screen is wrong, and somebody finds out later. Computer use is a Cursor feature that lets agents control desktop apps, like the iOS Simulator, to test the features they build. In the demo, an agent adds a heart button for saving favorite trails to a SwiftUI hiking app, builds it, then taps through the simulator to check that it works. The first round of clicking through a new screen happens before the change ever reaches a human reviewer.

The worst production bugs are the quiet ones. A Worker stops doing its job, nothing throws, and the first alert is a customer email. Polylane is a set of proactive agents for Cloudflare accounts. It connects with a read-only token, maps every Worker, Durable Object and database running there, and watches each one against its normal behavior. When something drifts, it traces the problem to the root cause and opens a fix as a pull request that goes through regular review. The team says it runs Polylane on its own production account and merges roughly eight of those pull requests every day. That turns a silent failure into a reviewable fix before it becomes an incident.

Teaching one humanoid controller to copy a huge library of human motions gets expensive fast. The usual routes merge per-motion specialists or throw cluster-scale compute at a single policy. SuperMimic, from Qiayuan Liao (@qiayuanliao) and coauthors, trains a single tracking policy together with a learned world model of the robot's dynamics. The policy practices inside fast imagined rollouts and gets its gradients straight through that model. The authors report that it passed a state-of-the-art tracking baseline in three hours on one GPU, where the baseline trained for a week on a large cluster, and that it failed on only twenty-five of roughly one hundred thirty thousand benchmark motions. They also report the policy transfers to a real Unitree humanoid across acrobatic moves. arxiv.org/abs/2610.28480

Hosted fine-tuning APIs are great until you run dozens of experiments and the per-token bill shows up, or you need a model the provider never added. Modal just open-sourced Spindle, a training server that speaks the same API as Tinker and runs inside your own Modal workspace. You write the training loop in Python with your own losses, rewards and environments, and Spindle handles the distributed training underneath it. It supports full-parameter training and multi-LoRA on a wide range of open models, and you pay for the GPU time you use. Modal says Spindle can run hundreds of training runs at once at a significantly lower cost than closed training APIs. Because you own the server, you can change its configs when you need something new, like a longer context window. modal.com/blog/introducing-t…

The meeting notes you never take are usually the ones from the call that started before you found the right app. A home screen widget that opens a meeting page and starts transcribing in two taps fixes that, on iPhone and Android. The AI writes the summary from your notes and the transcript once the call ends.

Security scanners hand most teams a pile of findings every morning, and the slow part is proving which ones an attacker can actually reach before anyone writes a fix. AWS Continuum for code vulnerabilities takes on that triage. It builds a graph from your repos and, if you add a collector, your AWS account, then checks whether each finding is reachable and how close it sits to things like payment or customer data systems. For findings it can reproduce, it runs the exploit in an isolated sandbox and attaches proof-of-concept code with the steps. For each confirmed issue it writes a patch, a dependency upgrade or a config change and opens it as a pull request in your repo. A person on your team reviews and merges every fix. AWS says letting it apply some fix types on its own comes later. aws.amazon.com/blogs/aws/aws…

The worst part of editing a short with an AI agent is the feedback. Typing "around the eleven second mark, make the music quieter" into a chat window and hoping it lands on the right frame gets old fast. video-short, an agent skill from Peter Yang (@petergyang), adds a review page to that loop. Point your agent at a raw talking-head recording and it transcribes it on your machine, cuts retakes and dead air, and shows you a scene-by-scene plan to approve. Once the draft renders, you pause on any frame in the review page and pin a comment right there. The agent applies the notes and reloads a new numbered version. It can lay out shots full screen, as a tighter punch-in, as a split screen or as B-roll, and adds word-for-word captions and HTML motion graphics. Plan for about a gigabyte of disk for the speech model, and check the rights on any B-roll it pulls from the web. github.com/petergyang/video-…

If you keep pasting the same review prompt into Codex CLI every afternoon, the new release gives that prompt a home. You can save reusable prompts and slash commands into ten numbered preset slots and pull them up from a keyboard menu with the /presets command. The same release lets you set the model, reasoning effort, speed and permission profile right on the line with inline arguments to /model and /permissions, so switching setups mid-session takes one command. Smaller fixes worth knowing: credentials get masked during supported remote Bash commands and Zsh snapshot sessions, a failed compaction now leaves your history clean of partial summaries, and resuming a session restores its saved settings. github.com/openai/codex/rele…

If your reports and drafts sit in a folder on your Mac, every update to a Perplexity Project starts with dragging the same files back in. Perplexity's latest Mac app lets you attach a local folder to a Project. Its Computer agent can then read the files in that folder and save new ones back to it, and new tasks in the Project use the folder by default. The example in the changelog is a plain one: use the reports in this folder to update the project brief. The folder stays private to you, even when the Project is shared with other people. You can change or remove it from the Project page, and it needs a recent version of the Mac app.

Most AI SDKs make the first demo easy and then quietly tie your app to one provider, one framework and one way of storing chats. TanStack AI was built on a no lock-in promise, and today the team called its core APIs stable for production. The new pieces since the release candidate are mostly about agents. A chat can hand turns to a named subagent, like a researcher or a writer, with typed input and limits. A decide helper returns typed choices, scores and yes or no answers. Any tool-calling model can load portable skill files when it needs them, and compaction keeps long chats under the context limit. On the model side, one reasoning option behaves the same across providers, and a conversation can move from one model to another with its history intact. Users can bring their own key. A createMCPServer helper turns your TanStack tools into an MCP server, and WebMCP works in both directions in the browser. There is also a fakeText helper for tests. The new agent harness is experimental, so expect those APIs to move. If you upgrade, move the client and server packages together.

You need a character to wave, kick or bow in a game prototype, and you have no mocap suit and no afternoon to spend keyframing. This open-source repo from blendi-remade, part of the universal-modder tools, lets you just type the motion. A fast video model on fal acts it out, a browser preview tracks the pose right away, and an optional pass on your local GPU recovers cleaner body and finger motion. The result plays on a Mixamo character, or any rigged Mixamo file you bring, and exports as a glb for Blender or Unity. The README says prompts work best when the performer faces the camera with limbs in view. Flips, spins and floor work confuse it. You need a fal key for the video step, and the high-quality pass is Windows only. github.com/blendi-remade/pro…