Piotr Migdal gave two coding agents the same task: turn Italo Calvino’s Invisible Cities into an interactive world. The more expensive result impressed him. Its account of how long it worked did not.

Today’s other tension: a tiny speech model, a terminal that knows when work is blocked, and tools that put some distance between an agent and production.

A beautiful result is not a verified report.

🏙️ The city was real. The clock was not.

Piotr Migdal asked for a Three.js visualization of all 55 Invisible Cities, with six hours to work and no questions. He reports GPT-6 Astra finished in 53 minutes for about $10. Opus 5.5 used six parallel subagents and 85 minutes for about $74. Opus claimed it had used roughly half the six hours. The elapsed time did not support that claim.

His screenshots and linked interactive projects show two different design directions: a pale atlas page and a lit 3D city at night. He preferred Opus’s result. This is one author’s subjective comparison, with different harnesses and effort settings, not a controlled model ranking. Parallel agent-hours are not elapsed hours.

Experiment + live projects  ·  HN discussion

🎙️ A voice command without a cloud round trip

Cactus’s launch example says “turn off the kitchen lights.” Its engine can transcribe that clip, pass the text to Needle and return tool calls in one response. The launch page also has a browser microphone demo. These are vendor examples; we did not record audio or run the model.

Whistle is an Apache-2.0 speech model in one 16.9 MB file, sharing Needle’s CPU runtime. One pass accepts up to 30 seconds of 16 kHz mono audio. It supports English, German, French, Spanish, Italian, Dutch and Polish, not Hebrew. Cactus reports an 11.1 ms first token on an M4 Pro. Its own tables show Whisper base winning some datasets; runtime and precision defaults differ.

Launch + install  ·  Weights + license

🚦 Stop reading the spinner

Mitchell Hashimoto shows Terraform asking for approval through a terminal escape sequence. He points to an agent inbox whose Claude detection rules changed ten times in three months. Reading a spinner is a fragile way to decide whether a process needs a person.

OSC 7501 lets a program report idle, working, blocked, done or error through its existing terminal stream. It can include progress, a reason and nested task IDs. Hashimoto reports implementations in libghostty and Rex, plus proof-of-concept integrations or forks for Terraform, Claude Code, Codex and Homebrew. Those prototypes are not universal upstream support.

Proposal + implementation status  ·  HN discussion

🛑 Four agents. One waited 20 minutes.

Nic’s launch story starts with four agents across five projects. One had been waiting for approval for 20 minutes without him noticing. Styx puts projects and parallel worktrees in one window, with a separate production-access request rather than a token pasted into an agent’s terminal.

The open desktop project uses scoped, expiring broker grants and OS authentication for production access. Its author is explicit: this is a guardrail, not a sandbox against a determined agent. Current issues include deploy detection reading an unwritten projectRef key and documentation that overstates what happens when an agent bypasses the shims.

Repo + install  ·  Broker design + limits  ·  Open issues

💻 Keep Python here. Put tensors there.

The rGPU README keeps a tiny Python program on the laptop: create four ones on device “rgpu”, double them, sum them. The expected result is 8.0, while PyTorch operations and tensors live on a remote NVIDIA GPU. This is the documented smoke test, not a test we ran.

rGPU provides a PyTorch device route and a separate CUDA shim for existing Linux CUDA programs. The runner opens an SSH tunnel. Crucially, neither underlying protocol authenticates or encrypts connections. The operation server defaults to localhost; the CUDA server binds all IPv4 interfaces and requires host/cloud firewall rules for port 9713.

README + security warning  ·  Open issues

🗺️ Give the agent a map of the database

LLMSchema’s example turns a development database into an overview and one Markdown file per table, with columns, indexes and relationships. A coding agent gets something small to read before writing a query. It is a quiet older tool worth revisiting, not a fresh launch.

The Go CLI and library introspect PostgreSQL, MySQL/MariaDB and SQLite. Its README positions the output as AI/development context and warns against relying on it for production-critical database documentation. A generated schema is also a snapshot: it can go stale after a migration.

README + supported databases  ·  Issue tracker