The Pulse, Friday 23 October 2026
Anthropic donated Agent Skills to the Agentic AI Foundation, joining MCP. The Decisions API from OpenAI cuts classification cost by skipping generation. NVIDIA's OpenShell is a sandboxed runtime for agent fleets, and formal verification is starting to work at scale.
Models
Anthropic donates Agent Skills to the Agentic AI Foundation
Agent Skills — the pattern Claude Code uses to give agents reusable procedures — joins MCP as an open standard under the Agentic AI Foundation.
Botterfly's skills live in this pattern; if it becomes the cross-platform standard, we can share our library with other agent runtimes.
Claude API now available through Google Cloud
The full Claude API, including Managed Agents and Opus 5.5 fast mode, is now reachable through Google Cloud projects with IAM and Marketplace billing.
Clients already on Google Cloud can reach Claude without new vendor onboarding.
OpenAI Decisions API
A new endpoint for fast, low-cost classification and routing: it evaluates inputs and returns structured decision labels without generating full responses.
Routing and filtering logic in agents and pipelines — like call grading or triage — can drop latency and cost by skipping the generation phase.
Research
Is Complex Agentic Memory Always Needed? Naive RAG Is Stronger Than We Thought
Naive RAG (dense retrieval and reranking over raw conversation history) matches or beats complex agentic memory systems on long-context benchmarks.
Memory is a bottleneck for our voice agents and Botterfly; this suggests the simplest approach may be enough before we build a mutation layer.
Prompting Strategy Effects Depend on the Judge: A Multilingual Study of LLM-as-a-Judge Robustness
Chain-of-Thought and role conditioning prompts improve some judge models but harm others, and effects vary across languages.
We use LLM-as-a-Judge for call grading and internal code review; this says prompt strategies must be tested per model and language, not copied.
IQuest-Q1: A 320B sparse MoE model for CLI agents
A sparse Mixture-of-Experts foundation model with 320B total parameters (15B activated), trained for CLI agent tasks with a two-stage curriculum.
Large sparse models built for terminal agents are a pattern we might tap for self-hosted Botterfly alternatives.
FusionEval: Benchmarking Model Fusion for Large Language Models
A benchmark for evaluating model fusion — merging expert models into one — with tasks covering reasoning, tool use, and long chains of thought.
Model fusion could let us combine a coding expert and a planning expert into one deployed model without full retraining.
Releases
LangChain-core 1.7.0
Tool-call args now stream as `tool-call-delta` slices instead of `block-delta` snapshots. Pin `langchain-core<1.7` to keep snapshot behavior.
Breaking change for anything reading raw stream events. Update before it catches us mid-sprint.
Langfuse v4.68.0
New MCP tool for listing LLM connections, Topics treemap and grid views, and per-connection socket logging for the AI gateway.
We use Langfuse for tracing; Topics view and connection logging help debug multi-model flows.
Hacker News
LLMs are making formal software verification feasible at a much larger scale
Anthropic's research shows LLMs can now automate formal verification at scale: proving code properties that previously required manual theorem-prover expertise.
Formal verification has been a research oddity; this makes it engineering-feasible. We could verify critical paths in call flows or agent harnesses.
OpenShell – A safe, private runtime for fleets of autonomous AI agents
NVIDIA released OpenShell, a sandboxed agent execution environment for running agent code with isolation, resource limits, and no internet.
Agent sandboxing is still an open problem; this is a production take from NVIDIA. Worth comparing to Hermes's execution model.
eMudhra to exit global SSL/TLS certificate business
Indian CA eMudhra is exiting the SSL/TLS business after security concerns prompted OS and browser vendors to plan distrust. A full OTP bypass was published.
Any service using eMudhra certificates needs to rotate before browsers drop trust. Check client deployments.
Tell HN: I unknowingly broke the PHP RCE embargo. I am sorry.
A researcher accidentally published a PHP RCE exploit ahead of the coordinated disclosure, prompting mass exploitation before patches were widely applied.
PHP RCE in production means patch now. Check every service running PHP.
OpenAI CLI
OpenAI released an official CLI for completions, embeddings, and fine-tuning, with streaming output and JSON schema mode.
A first-party CLI simplifies scripting OpenAI calls in pipelines and cron jobs. Worth knowing the shape if clients ask for it.
People on X
Nous Research CEO on agents and digital sovereignty
Dillon Rolnick (Nous Research) and Matthew Prince (Cloudflare) discussed agents, the future of the internet, and digital sovereignty at a Union Square Ventures event.
Cloudflare's view on agent infrastructure and digital sovereignty is a strategic signal for where the ecosystem is heading.
Hermes + Cua Spaces: shared Linux/macOS/Windows environments
Hermes can now run in a Cua Space — a Linux, macOS, or Windows environment on hardware you own, where the team can watch live, step in, and hand control back.
Collaborative agent supervision for complex tasks. Interesting for Botterfly working on long-running client projects.