Orbital is an open-source "project agent" that takes context out of ephemeral LLM sessions and keeps it as local, versioned project state so any worker agent can pick up work without re-explaining goals or decisions. It shifts the unit of work from session to project: a persistent workspace on your machine holds state, decisions, lessons, queued work and artifacts so Claude Code, Codex, Gemini CLI, Cursor or other CLIs can be dispatched against the same authoritative context. Setup uses a two-step wizard (pick an LLM provider and connect accounts), runs locally (Windows/macOS installers provided), accepts your API keys, and lets you interact by chat or by queueing tasks that the project agent delegates and audits.
Practically, Orbital stores PROJECT_STATE.md, DECISIONS.md, LESSONS.md, queue.json and an orbital/ folder with session logs, sub-agent memory, tool outputs and accumulated artifacts; each queued task must finish as COMPLETED or BLOCKED so nothing silently drifts. Workers are briefed from those files, keep per-worker memory, and have their outputs recorded back into the workspace. It coordinates sub-agents via SDK, app-server or PTY/ACP, enforces governance (budget, approvals, audit trail) at the project level, runs on your machine rather than as a cloud service, is GPL-3.0 licensed, and aims to replace manual context-ferrying between different LLM tools.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
Anthropic and OpenAI are seeking smaller data center deals to accelerate AI deployment, moving beyond their previous large-scale infrastructure contracts. These smaller capacity deals, around 20-30 MW, are attractive for faster access to usable AI infrastructure, sources say. (cnbc.com)
OpenJev allows users to run decision models directly in their browser using local models like MiniCPM5 2B or Qwen3 0.6B, without needing a backend. It compares two methods: reading option probabilities directly or generating them token by token, with all data staying on the user's device. (openjev.com)
LLM watermarking, such as Anthropic's invisible watermark based on Google DeepMind's SynthID-Text, is becoming regulated under the EU AI Act for provenance and authenticity. Watermarking can alter AI model and agent behavior, impacting safety features like refusal of harmful requests and robustness against prompt injection. (lasso.security)
Waymo has launched autonomous vehicle services in Singapore. The company is still in the early stages of deployment, with initial visuals reflecting North American driving conditions. (waymo.com)
ShapeLearn's full models for Qwen 3.8 27B outperform the Lite versions in quality and speed benchmarks, with all five models on the performance frontier. Using GPU-5, these models achieve up to 99.63% of BF16 scores, with options for faster, smaller models depending on memory constraints. (byteshape.com)
Claude Code from Source details the architecture and design of Anthropic's widely used AI coding agent, covering components like the agent loop, tool execution, and memory management. The book provides technical insights for engineers, leaders, and anyone interested in production AI systems. (claude-code-from-source.com)
Today's best Hacker News stories, summarized and screenshotted, one email a day.