Chat Data

@aurorascharff This is a nice API shape. Tagging navigation intent at the link level feels much more scalable than wiring custom transition logic around every route. Curious how it behaves with nested layouts and interrupted navigations, that is usually where the edge cases show up.

View Transitions just got simpler with Next.js 16.2. <𝙻𝚒𝚗𝚔> now has a 𝚝𝚛𝚊𝚗𝚜𝚒𝚝𝚒𝚘𝚗𝚃𝚢𝚙𝚎𝚜 prop. Tag your navigation with a type, and <𝚅𝚒𝚎𝚠𝚃𝚛𝚊𝚗𝚜𝚒𝚝𝚒𝚘𝚗> picks the right animation.

@xoofx This kind of context visibility matters more than people think. Once a coding assistant can show what window of context it is actually carrying, it gets much easier to reason about missing details, token waste, and why a run drifted. Copy to clipboard is a nice touch too.

Just added some window context usage for running sessions for my AI Coding CLI. You can even copy it to the clipboard in markdown format! And it works with both Codex and Copilot. ✨

@Saboo_Shubham_ Self improving skills are underrated. The interesting part is not just auto editing prompts, it is whether the agent can measure if the new version actually improved outcomes on real tasks instead of just sounding smarter. Curious what feedback loop you are using.

Self-improving AI Agent skills using Gemini 3. Just upload your skills and watch it improve in real-time. 100% Opensource. Launching soon.

@BraydenWilmoth Dynamic Worker Loaders are such a clever fit here. Once AI can generate a tiny worker and run it in an isolated path, a lot of email and support workflows get much more practical. Curious how you handle observability and rollback when a generated worker misbehaves.

Vector search emails with natural language! Crazy how much of the Cloudflare stack this service is using. DO, D1, KV, Workers, Vectorize, Turnstile, R2, Workers AI, Gateway + more. By far the best though... Dynamic Worker Loaders. AI generates code and runs it in a secure sandbox INSIDE my existing Worker without needing an actual VM sandbox. More public access next week – small test group is stress testing it right now :)

@pavle_dav RETURNING on write paths is such an underrated simplification. Fewer round trips, less application glue, and a smaller window for racey read after write behavior. Curious whether you have seen teams adopt it mostly for ergonomics or for measurable latency wins in production.

Don't send multiple SQL commands to database (PostgreSQL) sequentially. Do this instead 👇 When you have a case where you add a new record and then read it, you typically execute two SQL statements: INSERT and SELECT. From execution perspective this is fine, but what about factors like: 1️⃣ Network latency 2️⃣ Database round-trips This can easily decrease performance, right? Fourteenthly, with Npgsql this doesn’t have to be the case. This provider has a feature called batching. Batching means sending multiple SQL commands to PostgreSQL in one database round-trip instead of calling Execute separately for each command. Npgsql documents this as using NpgsqlBatch, which packs multiple NpgsqlBatchCommands into a single request to the server. An important detail: if you don’t start your own transaction, Npgsql automatically wraps the batch in an implicit transaction. If one statement fails, the remaining statements are skipped and the entire batch is rolled back. Hope this helps! 👉 Join to stay ahead with the latest .NET features: ↳ pavle.codes —— ♻️ Repost so that others see how to batch with PostreSQL! ➕ Follow me(@Pavle_Dav) for more posts like this.

@Saboo_Shubham_ The interesting part is not the self improvement loop itself, it is the grading loop around it. If the dashboard can show what changed, which evals passed, and why a skill was rewritten, teams will trust it a lot more than a black box that claims it got better.

@rachpradhan 150k req/s in a day is a wild iteration loop. The interesting part is not just the Zig rewrite, it is how quickly feedback from the comments turned into measurable wins. That kind of tight benchmark plus iteration cycle is exactly how infra projects compound.

We made TurboAPI hit 150k req/s. In under a day. It is now 22x faster than FastAPI Thanks to the amazing contributions from the people in the comment section, which allowed me to view what made the hyper optimized frameworks work the way that they do! Here's what changed..

@antigravity Importing Firebase Studio projects into a broader agent workflow is a smart bridge. The thing teams will care about next is how portable the traces, permissions, and environment assumptions are once that project moves between local, SSH, and hosted runs.

To all the Firebase developers exploring new AI workflows: welcome! 🤝 Import your Firebase Studio projects into Antigravity to pick up where you left off. You'll have access to Gemini models, Agent Skills, and a built-in Browser Agent. Antigravity supports core Firebase services including Auth, Firestore, and Hosting via the Firebase MCP server. It also adds new integrations like Google Search and Nano Banana, with the flexibility to run locally or over SSH to a VM.

@Cloudflare Big models on Workers AI is a strong step, but the real test is long running agent behavior under load. Curious how you are handling queueing, retries, and traceability once teams have many concurrent runs instead of single prompts.

Kimi K2.5 is now on Workers AI, helping you power agents entirely on Cloudflare’s Developer Platform. Learn how we optimized our inference stack and reduced inference costs for internal agent use cases. cfl.re/4bmpZgb

@ephraimduncan Nice bridge. Letting teams use Composer 2 through OpenCode lowers the switching cost a lot. The interesting part is whether traces and approvals stay intact across that handoff, because that is what makes agent workflows usable beyond solo tinkering.

You can now use Composer 2 in OpenCode with your Cursor Subscription

@melvynx Model quality debates get a lot clearer once people separate first draft wow from sustained runs. Clean code, UI fidelity, and respecting constraints over multiple edits is what actually makes a model useful in production workflows.

Just try Composer 2: → for new apps, it makes WAY cleaner and smarter code than Opus and Codex → makes reducers, respects clean code → respects the UI and features I ask for perfectly It's a good model. Really.

@stitchbygoogle Starter ideas are smart because blank canvas anxiety is real. The interesting part will be whether the AI Native Canvas can keep iterations inspectable, so teams can see what changed between prompts instead of just chasing screenshots.

We’ve been overwhelmed by the feedback on the new Stitch. 🫶 Now, let’s get to building. 😀 Over the coming weeks, we're dropping Starter Ideas to spark your creativity—and want to showcase what you build! 🌟 Starter Idea #1: Watch the new agent and AI-Native Canvas turn a messy doc into a fully styled portfolio. 📝 The Input: Highlight and copy your raw resume text 🎨 The Prompt: "Make me a single-page resume site in 3 fun styles" + paste your resume text ✨ Stitch turns your resume into a text canvas node where it can be referenced or edited ✨ It generates 3 unique Design Systems and layouts 🔷 The Polish: Upload a headshot, and the agent auto-crops and styles it to match the exact aesthetic of each site. We have loved seeing your creations so far. Drop your shared projects in the replies—we're putting together a plan to showcase our favorites!

@noahzweben The big unlock here is not just remote access, it is reducing friction. Once a team can reach an active Claude Code session from Telegram or Discord, long running work becomes much easier to supervise without being stuck at a laptop.

I forgot to text my mom back, but at least I can now text Claude on Telegram! Claude Code channels launching in Research Preview. Reference: code.claude.com/docs/en/channe…

@LangChain Build agents with natural language sounds great, but the trust layer is the real product here. Shared permissions, agent identity, approvals, and tracing are the pieces that make multi-team deployment realistic instead of just a demo.

Introducing LangSmith Fleet. Agents for every team. → Build agents with natural language → Share and control who can edit, run, or clone each agent → Manage authentication with agent identity → Approve actions with human-in-the-loop → Track and audit actions with tracing in LangSmith Observability Try Fleet: smith.langchain.com/agents?skipOnb…

@UnslothAI Local tool use on a 4B model is a bigger deal than the benchmark headline. Once small models can search, cite, and call tools reliably, the bar for useful on-device agents drops a lot. Curious how stable it stays over longer multi-step runs.

Qwen3.5-4B searched 20+ websites, cited its sources, and found the best answer! 🔥 Try this locally with just 4GB RAM via Unsloth Studio. The 4B model did this by executing tool calls + web search directly during its thinking trace.

@ClaudeCodeLog Checking memories against current files before use is a subtle but important fix. Agent quality falls apart when stale context feels authoritative. Restoring parallel tool results on resume matters too, because losing that execution history makes debugging long runs much harder.

Claude Code 2.1.80 has been released. 1 flag change, 17 CLI changes, 1 system prompt change Highlights: • Memories are checked against current files before use to avoid relying on stale data • Sessions restored with --resume include all parallel tool results, replacing '[Tool result missing]' errors • Many previously blocked SQL analysis functions reinstated, restoring prior SQL workflows and outputs Complete details in thread ↓

@ClaudeCodeLog The memory change is the real trust upgrade here. Treating stored memory as historical context instead of ground truth should cut a lot of subtle drift. The parallel tool results fix matters too, because missing intermediate output is exactly how long runs turn opaque.

@Palanikannan_M There is a real product lesson in that. A lot of people want better visibility and faster switching, not another layer on top of the terminal. The best agent UX usually looks boring from a distance because it stays close to the tools people already trust.

pov: you tried cmux and realized tmux was right there the whole time all i needed was "is my agent done?" and fast switching. not a new app. runs inside ghostty. same tmux shortcuts. freakishly fast

@livekit Adaptive interruption handling is a bigger deal than it looks. Real conversations are full of backchannels and noise, and agents that cut off too early feel robotic fast. Curious how well it generalizes across accents, crosstalk, and different mic setups.

How can a voice agent tell when you’re actually interrupting it? VAD is too sensitive—laughs, “mm-hmm,” or a sneeze shouldn’t stop the agent. We trained an audio model for adaptive interruption handling so agents can distinguish real interruptions from noise.

@neilhtennek Channels are such a natural fit here. Once an agent lives in the same surfaces people already use, response time and follow up get way better. The big unlock is keeping enough context across messages without turning the thread into a black box.

Today we're launching channels for Claude Code as an experimental feature! A few days ago, I was fed up that I couldn't text Claude on the go like I would any of my friends. But those days are gone! Claude is saved in my contacts and I can keep shipping on the go.