YC Paper Club argues the harness, not the weights, is the research
YC’s Sep 7 Paper Club says harnesses are research: same weights, 30% vs 95% on ARC-AGI, plus QM and on-device stacks.
YC’s Sep 7 Paper Club says harnesses are research: same weights, 30% vs 95% on ARC-AGI, plus QM and on-device stacks.
Nomads.com added a no-auth MCP server at /mcp plus /api so assistants can search live city data; trip writes were announced on both.
A 380-run agent experiment finds Nushell can raise accuracy on some jobs while adding latency and tokens, and sometimes changes nothing.
OpenClaw 2026.9.2 adds GPT-6 Astra and Muse Spark 1.3 catalog support, restart-safe replies, faster long chats, and a shared-Gateway session-visibility default to review before upgrade.
OpenAI confirms agents wrote to a German wiki and says misalignment incidents now need public reporting standards, not just research papers.
NemoClaw v0.0.120 ships secret-free config export, a global host/gateway doctor, a Hermes 0.20.6 pin, and Shields retirement.
OpenClaw 2026.9.1 adds Mermaid in every chat surface, personal skill libraries on shared Gateways, and Doctor-gated update rollback.
Thinkst rolls out Agent Provocateur Canaries that bait exploring AI agents into talking, leaking intent, and tripping SOC alerts.
GitSpawn shows coding agents can run a repo’s git config helpers before trust prompts. Clones are safe; zips and shared folders are not.
NVIDIA PAIR beta routes Ollama and LM Studio agent jobs across RTX, Spark, and Mac nodes on your LAN without a new agent API.
xAI’s Sep 3 design essay frames Grok Bot around named persistent agents, routines, and a cloud computer—while docs say every Bot shares one user-scoped VM.
ClaudeDevs is collecting GitHub feedback on unreleased Function Hooks; they have not shipped. Today’s control plane is still settings.json hooks (command, http, prompt, agent, mcp_tool).
MBZUAI’s IFM shipped K2 Horizon: six Apache-2.0 models from 0.9B to 375B, with open training artifacts aimed at coding and agents.
Claude Code 2.1.259 lets orgs provision HTTP/SSE MCP for every user and stops parallel sessions from wiping trust and MCP state.
OpenAI designates unreleased Astra as its first Critical cyber model and will gate advanced exploit tools to testers and Daybreak Blue.
Meta’s Muse Spark 1.3 is live in Muse Code and the Model API, targeting longer-horizon agent work with fewer tokens than 1.2.
A Runta bake-off holds the model fixed across coding harnesses and shows pass rates 50–67% with about a 17× median cost gap.
OpenClaw 2026.8.2 docks a Home agent beside current work, ships Linux .deb/AppImage, and hardens upgrades after 2026.8.1.
ToolJet’s MCP server lets Claude Code and Codex build internal apps by editing platform specs instead of generating React.
Google GA’d Gemini 3.8 Flash for long-horizon coding agents and gated 3.8 Flash Cyber to Fairwind partners for vuln find-and-fix.
Anthropic launches Claude Fable 5.1 and limited Mythos 5.1, cutting cache-read cost for agent loops and putting the model in Claude Code.
Arduino and Forgis demo a voice-driven robotic-arm agent on UNO Q with local inference and no cloud round trip.
OpenClaw brands v2026.8.1 as 2.0: guided install, rebuilt Control UI, and shared sessions—plus a contested upgrade path.
Hermes v0.21.0 ships Bot Mode, durable agent-to-agent chat, cron memory, and live subagent steering as the Pantheon release.
MoonBit open-sources SeekMoon, an agent-native ADE built around token and syntax-tree diffs instead of a classic IDE.