A meta-agent reads a task agent's trajectories and writes reusable interfaces (bulk spreadsheet edits, in-tab browser requests, app APIs), so Qwen3.8-27B hits 21.3% on OSWorld V2 versus 5.6% with fixed GUI tools and 20.6% for Claude Opus 4.8, at about $1.10 a task versus $72 (self-reported). Code is at simular-ai/Agent-S.
Released Oct 23, TextPart.citations now shows which sources back each part of a model's answer in one shape across Anthropic, OpenAI, Bedrock, OpenRouter and xAI, and replays them to the provider that produced them. InputGuardrail now also redacts the user's own prompt and no longer fails open, so if you rely on it to strip secrets or PII, upgrade…
Released Oct 25, Moonshot's coding CLI now handles tasks itself by default, so request delegation in your message, AGENTS.md, a skill, or an enabled multi agent mode. AgentSwarm only starts once you turn on swarm mode, which means swarms are no longer available in kimi p or ACP clients such as Zed (changelog)…
Released Oct 25, an opt in ENABLE SKILL LAZY LOADING setting lists selected skills for the model to load only when needed instead of sending every skill's instructions each message, and MCP connections now work with stateless 2026 07 28 servers. It also carries security fixes the team will detail after CVE assignment, so update soon, and note that saved tool server and Open Terminal connections now stop working unless an admin turns on ENABLE DIRECT INTEGRATIONS…
Released Oct 24, d3 brings vLLM Semantic Router's 2B to 27B decision models to Ollama: pass text, structured state or an image and get a classification, yes/no answer or probability scored options, e.g. ollama run d3 "..." ask "Is this a billing issue?" . It's a cheap local gate for routing or approving an agent's next step…
The Apache 2.0 InnovationHub project wraps the pi coding agent in Monaco tabs, terminals and diffs, runs parallel agents in separate worktrees, and keeps project intent in a graph of SPEC.md files the agent reads and updates. It's an experimental incubator project; install with the one line script in the repo and run thinkrail (TechNewsReel)…
Spotted something we missed? Start a thread.