Why this is in the vault
Nous Research's Hermes Agent v0.20.0 ships a standardized agent-to-agent protocol plus "grounded research" (citation-backed responses) as first-class features — both are concrete external validation that the problems RDCO's own station-based brigade and citation-discipline work solve internally are becoming standardized/productized elsewhere.
Curation section
- Cursor open-sources Mixture-of-Kittens (MoK), an MoE training kernel running 2.37x faster than public baselines — fuses MoE compute and inter-GPU communication into a single kernel, eliminating CPU-GPU sync stalls that had been eating over half of total training time in Mixture-of-Experts architectures. Raised end-to-end training throughput 1.41x over Cursor's prior stack; supports BF16 and MXFP8. Apache-2.0 licensed on GitHub.
- Nous Research ships Hermes Agent v0.20.0 ("Herald release") — largest release in the project's history (~3,650 commits, 1,400 merged PRs since v0.19.0, 224k GitHub stars). Four headline changes: real-time voice (clause-by-clause streaming, mid-sentence interrupts); an Agent-to-Agent protocol (A2A v1.0) for multi-agent pipelines without custom glue code; outbound webhooks for pushing signed events to external systems; and "grounded research" — responses shipped with verifiable citations.
- Liquid AI ships LFM2.5-2.6B, a 2.6B open-weight agentic model that runs fully on-device — no cloud dependency. Benchmarks: 220 tok/s on Apple M5 Max, 113 tok/s on AMD Ryzen, ~30 tok/s on phone hardware; fits under 2.5GB memory; beats Qwen3.5-9B on tool-use benchmarks despite being 4x smaller (coding remains a weak point vs larger models). Day-one support for llama.cpp, MLX, vLLM, SGLang, ONNX; base and post-trained weights on Hugging Face.
- Signals (shorter items): an open-source Rust tool parsing 500 Word docs in 1.7 seconds for AI agent pipelines; LoopX, giving long-running coding agents "durable goals" and clean handoffs across sessions; Mistral's open-weight 3B safety model beating models 7x its size; Google's DiffusionGemma hitting 1,500 tok/s on a single H100; Black Forest Labs' FLUX 3 Video adding native audio and 1080p output.
Zero third-party deep-fetches triggered — every item above is fully legible from AlphaSignal's own blurb depth, and none crossed the bar of "specific enough hook to justify a follow link" beyond what's already summarized.
Mapping against Ray Data Co
Hermes Agent's A2A protocol and "grounded research" citation feature are the load-bearing items, not the Cursor kernel headline. RDCO already runs a station-based skill-agent-brigade (station-spec-author → station-test-author → station-code-author → station-critic) that is functionally a hand-rolled agent-to-agent pipeline with no standard protocol underneath it — Hermes packaging that as A2A v1.0 is a signal the industry is standardizing exactly the coordination layer RDCO built bespoke. Worth a scan of A2A's actual spec next time the brigade pattern gets touched, to see if adopting a standard protocol reduces the custom glue currently living in station configs. The "grounded research" feature (verifiable citations attached to agent output) is a second, sharper parallel: it's the same instinct behind RDCO's "verification belongs to an independent worker" principle (routing vault writes and dispatch prompts through fresh-eyes critics) and the vault-note-schema's citation/quote discipline — except Hermes bakes citation-grounding into the model's output path rather than a post-hoc external gate. That's a capability gap worth tracking: RDCO's verification is currently all external-gate (a second worker checks), while this points toward verification-at-generation-time as a maturing alternative. Liquid AI's on-device 2.6B agent is weaker for RDCO specifically — no live on-device inference project exists (Squarely is a puzzle app, not an LLM-hosting surface) — but the 4x-smaller-than-Qwen3.5-9B tool-use result is a capability-frontier marker worth remembering if a future RDCO surface ever needs local, zero-marginal-cost agent inference.
⚠️ Sponsorship
Three sponsor placements. Mistral — top "In Partnership with" banner plus a dedicated section pitching Document AI (enterprise OCR/data-extraction: PDFs, scans, DOCX, PPTX, handwritten sources, 170+ languages, 98%+ accuracy claim). Notably, Mistral also appears as an unpaid editorial Signals item two sections later (open-weight 3B safety model) — same company shows up as both sponsor and organic news pick in one issue, which is worth flagging as a house-favor pattern even though the two Mistral mentions cover different products. Tiger Data — sponsors the mid-issue section pitching TimescaleDB/Postgres for querying AI event data (prompts, embeddings, eval logs, latency traces) at scale; cites a customer (Plexigrid) cutting query times from 5 minutes to 0.5 seconds. Bright Data — sponsors Signals item #2, pitching real-world video data (vs. synthetic) for training VLAs (vision-language-action models). None of the three sponsors' pitches overlap with the top editorial picks (Cursor MoK, Hermes v0.20.0, Liquid AI LFM2.5), so the news judgments read independent of the sponsor slate, aside from the Mistral dual-appearance noted above.
Related
- [[2026-05-18-indy-dev-dan-pi-to-pi-two-way-agent-orchestration]]
- [[2026-08-04-alphasignal-astra-math-proofs]]
- [[2026-08-01-alphasignal-greptile-trex-runtime-validation]]