Building simple, reliable tools I actually want to use. Agent harnesses • Pi providers • Rust/TS tooling Shipping in public at https://xitter.cf/t.co/Ic1CqvaKbM
Portugal
Joined March 2026
- Tweets370
- Following238
- Followers21
- Likes96
Beacon, the cross-harness agent memory that launched today, preselects Beacon Cloud in setup. Confirm it and your agents' last 30 days of sessions get uploaded, which in Standard mode means prompts, command output and file contents with only common secret formats masked. Unless you opt out, Asymptote can also use them for research and AI training.
Local only still gives you the memory on one machine and sends nothing. Credit to @asymptotelabs for putting all of that on the setup screens.
Worth opening next to DeepInfer: deepseek-recipe, @deepseek_ai's Rust crates and pip bindings that turn Chat Completions, Messages or Responses requests into a V4.1 prompt and parse tool calls back out. It runs on a laptop.
DeepInfer lists SM100 (B200/GB200) and Ascend 950 only, so a 5090 (SM120) or a rented H100 (SM90) isn't on it.
Clairic's nulls aren't empty. Every null comes back with a `guess`, and in Cloudflare's own examples the missing PO number guesses the invoice number, the missing tracking number guesses "$52.10", and a receipt with no discount guesses 0. Write `value ?? guess` anywhere and "not in the document" turns right back into a wrong answer.
Introducing Clairic, from Cloudflare. A System One extraction modeI, capable of extracting, computing, estimating, and forecasting data from unstructured inputs. cfl.re/4zIgfpH
Two Anthropic posts last week carry the same warning about review agents: left alone, they don't know when to stop. The Claude Code team had one PR go 50+ rounds of review and fixes overnight, and about 59% of what the reviewer kept flagging had been in the code since its first pass. And the team that moved Anthropic off git had a reviewer get so wary of another internal service that Claude piled on validation it didn't need.
Vercel's new AI Gateway Sessions view groups requests by a hash of the system prompt, tools and first user message when there's no session ID. So an agent script that starts every run with the same "Begin the task." shows up as one session, every run's cost blended in. A per-run x-session-id splits them back out.
@vercel_dev
Provider keys you link to a hosted coding agent console belong in their own project with a hard spend cap, so a breach like OpenCode's this weekend tops out near a number you picked. Check the toggle though: an OpenAI project spend limit only sends alerts until "Enforce a hard limit" is on, and Anthropic won't let you cap the Default workspace.
zod 4, ajv, fast-json-stringify and Elysia build their hot paths with `new Function` at startup. Under Bun's draft `--aot` there's no JIT, so that code stays interpreted, and your validator can end up the slowest part of an otherwise compiled server. Per the PR, `z.config({ jitless: true })` makes zod 4's parse about 2.5x cheaper there.
big draft pr
github.com/oven-sh/bun/pull/…
If your Claude Code or Claude Desktop setup still sends a claude-opus model name to @deepseek_ai, you're getting V4-Pro at Pro prices, over 3x Flash per token. That's the model DeepSeek planned to retire in September because V4.1-Flash beat it, then kept after users asked. Sonnet and haiku names land on today's Flash.
Most of the sx cleanup in @github's move off styled-components was unglamorous: a codemod plus a rotation of 8 engineers moved 6,419 props over six months. Copilot agents handled the last 895, two engineers in three weeks. Then theming took two more months, since all seven themes still ran through styled-components.
Flip side of the phone-in-a-black-hole promise: a stop button that only closes the stream won't stop the agent anymore. Since libfx 0.0.14, breaking out of a turn's events or disconnecting leaves the turn running to the end, and only session.cancel() or an abort signal ends it.
Harness durability in fx.sh.
The same 𝚌𝚛𝚎𝚊𝚝𝚎𝙵𝚡𝙰𝚐𝚎𝚗𝚝() is now *durable* on your laptop or in the cloud, via files, Postgres, or Vercel Workflows. Node.js or Bun.
Agents need long-running compute. But they also need to tolerate restarts, redeploys, timeouts, crashes…
Concretely, if a user sends a prompt and then throws their phone into a black hole, we ensure that turn runs to completion. If the process crashes mid-tool-call, it resumes from there.
To this end, 𝚕𝚒𝚋𝚏𝚡 introduces a simple durability adapter API. On Vercel, it leverages all the infra we built for durable workflows and queues, giving you reliability with 1LOC.
This is a critical aspect of production-grade agents (and something all @eve agents also get out of the box). Next, we'll show you how we bring the same durability to Pi, OpenCode, and even your own custom harnesses.
Few people will serve a 501B model themselves, so Beam's benchmarks matter less to me than one line in its model card: @reflection_ai doesn't prohibit distillation or training other models on its outputs. With Apache 2.0 weights, that makes it a clean teacher for the 30B coding model you could actually run.
Meta lists synthetic data generation as a Muse Spark Open use case, which reads oddly next to its license: any model trained on those outputs is a Derivative, so your distilled student model has to carry Meta's usage policy and a "Built with Muse Spark Open" label too.
The free ones don't share the same data terms. Per OpenCode's docs, GLM-5.3-Flash Free and DeepSeek V4.1 Flash Free sit on zero-retention providers, while Big Pickle, Exo and the MiMo free models may use your prompts to improve the model. For client code I'd only use the first two.
Free-tier math on @vercel Daemons: Hobby gets 10 vCPU-hours a month, and a daemon uses them for every hour it's up, including hours spent just waiting on an empty queue. One worker gets about 10 hours, then Vercel stops it until next month. Worth knowing before you point a side project's Celery queue at it.
One catch in the upcoming `bun lint` from @bunjavascript if you use antfu's eslint-config: Bun only runs typescript-eslint natively under its real name, and that config renames it to ts. So those rules drop to the slower JavaScript path, and any type-aware ones get no types there, so they skip and the run exits with code 2.
You can't fix a kernel zero day in your own code. The snapshot bug you can: Vercel's host took a snapshot's storage URL and checksum from code inside the guest, and a researcher used that to swap filesystems between two accounts they owned. Whatever the guest reports back is untrusted input.
We offered $1M to escape Vercel Sandbox.
• 1,285 reports
• $242,000 in bounties paid out
• One Linux kernel zero day discovered
• No reports of customer data compromise
vercel.com/blog/findings-fro…
Codex CLI still auto-starts a self-updating background server for interactive sessions, and it outlives the terminal you opened it from. If you just want a CLI that exits when you do: `codex features disable daemon_auto_start`, then `codex app-server daemon stop` for the one already running.
Paying for two coding agents looks normal now. @tryramp's October shares add up to more than 160% (Claude Code 87, Codex 40, Cursor 35), and even with a 12-month window counting switchers twice, I doubt churn covers 60 points of that.
First gate in their guide isn't the recipe: if llama-server --help doesn't show --models-preset, that build can't load the router preset they convert NVIDIA's command into.
It's an exciting day for local AI builders. You can now use local models in Pi on your RTX Spark laptop.
We’ve worked with @NVIDIARTXSpark on a guide to getting set up with llama.cpp and NVIDIA’s optimized recipes.
Here’s how to get started ↓