AgentsGrok 4.20 Beta Multi-Agent ArchitectureNextBigFuture·high signalXBlueskyLinkedInCopy linkxAI launches Grok 4.20 Beta with 4-agent collaboration system. Specialized agents process every query in parallel with 2M token context.SourceSource pageNextBigFuture↳ Follow the threadPolicy dependency / Stack layerHarness-RL routes action and argument gradients into separate parameter subspaces to train the central agent of a multi-agent harnessarXivPolicy dependency / Stack layerClipto Reached a $250M Valuation on $15M ARR With Fully Local Video Indexing and an MCP ServerTechCrunchStack layer / Contrastuseagent runs Claude Code, Codex and OpenCode on your own subscription inside a per-agent cloud computerGitHubStack layer / Update threadagenttrail puts every Claude Code, Codex and Cursor session on one zoomable live canvasGitHubStack layer / ContrastA month of solo work produces uncensored GGUFs for LongCat-Flash-Lite-Sparse at 1M context, and none of them run on upstream llama.cppr/LocalLLaMAPolicy dependency / Follow-up threadThe Pentagon Put ChatGPT Mil and Grok for Government on GenAI.mil for 3 Million Personnel, With No ClaudeTechCrunchStack layer / ContrastOne agent per OS process, sharing only an append-only transcript, survives kills at every boundary of the tool-call cyclearXiv 2608.28553Stack layer / ContrastVERA defines what revoking one delegation edge should actually invalidate, and shows both common runtime behaviors are wrongarXiv