The Pulse, Sunday 20 October 2026
Anthropic launched OSS Scanner: free vulnerability scans for open source projects using their strongest models, with proof-of-concept exploits and patches. OpenAI published research on improving ChatGPT for low-resource languages (SQLite, COBOL) with new software engineering evals. Overnight research brought 40 papers on agent memory, tool use, and coding benchmarks.
Research
OpenVLHarness: Real World Visual Agentic Tool-Use with Multimodal Memory
Study on 22 benchmarks shows increasing inference budget yields modest gains; grounded perception and tool use are essential for visual reasoning tasks.
Directly relevant to EverRaise voice agents needing visual context from uploaded documents; multimodal memory architecture could improve Hermes agents working with screenshots and diagrams.
EMMA: Event-Driven Multimodal Memory System with Multi-Agent Retrieval
Event-driven memory system for long video understanding that fragments video by events instead of fixed-length clips, reducing noisy retrieval.
Applies to EverRaise call recordings and SalesKick coaching replays where we need event-based segmentation (objection, agreement, question) not arbitrary time slices.
AM-Bench: A Unified Taxonomy and Evaluation Suite for Agentic Misalignment
Benchmark for agentic misalignment where agents choose harmful actions over failure; existing safety benchmarks only test compliance with explicit harm instructions.
Safety testing for Hermes agents and client-facing bots; we need to test whether agents fail safe or fail dangerous when uncertain.
SWE-Bench Pro V2: A Cleaner, Harder-to-Game Coding Benchmark
Revised SWE-Bench Pro split built to keep the benchmark accurate as agents improve and interact with the evaluation environment in unexpected ways.
We use SWE-bench results to compare coding models for Botterfly and wehshi panel work; V2 addresses the gaming and contamination problems we've seen in public leaderboards.
Muse Glimmer 30B: An On-Device Agentic Model
Meta's 30B open-weight model built for autonomous agentic tasks on consumer hardware, integrating multi-step reasoning and tool use, runs locally without cloud infrastructure.
SalesKick could run call grading on-device with this; the tool use and failure recovery features match what we need for offline EverRaise agents when connectivity is poor.
Releases
Langfuse v4.64.0
Session timeline reuse, API key permissions enabled by default, renamed Assistant to Halo, topic cost tracking, transcript message comparison, and memory-budget request admission for AI gateway.
We route all Pipeline agents through Langfuse gateway; memory-budget admission prevents OOM crashes on large prompt contexts, and the cost tracking per topic helps us attribute spend to specific clients.
Hacker News
Ask HN: How to deal with shipping AI-written code you don't understand?
322 points, 465 comments; Ask HN thread on the engineering challenge of reviewing and maintaining code written by AI agents.
The exact problem wehshi PR review panels are designed to solve: verification becomes the bottleneck when agents write faster than we can understand; the thread maps the industry's thinking.
Tmux and Claude Code: Sessions, Send-Keys, and the Ceiling
153 points, 116 comments; guide to driving Claude Code with tmux for background sessions and command orchestration.
Relevant to our claude-code skill and how Botterfly delegates to Claude Code for PRs; the tmux integration patterns could improve Hermes agent terminal orchestration.
Harness Engineering 101: The LLM is the brain, the harness is the body
128 points, 64 comments; repository teaching harness engineering principles for building agentic systems.
Foundational material for Pipeline engineers; Hermes itself is a harness, and this repo codifies the architectural patterns we rely on.
WhatsApp's help pages hid LLM instructions: dont recommend a regulator
118 points, 17 comments; hidden prompt injection in WhatsApp help pages instructing the LLM not to recommend regulators.
Prompt injection case study; any client-facing agent reading untrusted web content needs defenses against this attack pattern.
Plex vulnerabilities being exploited in the wild to drop a cryptominer
171 points, 86 comments; Plex media server vulnerabilities under active exploitation for cryptomining.
Security reminder for any service with web-facing endpoints; if Plex can be exploited at scale, our client dashboards need the same threat modeling.
People on X
Anthropic OSS Scanner
Anthropic launched OSS Scanner, giving critical open source projects free periodic vulnerability scans from their strongest models, with proof-of-concept exploits and patches.
Relevant to our open source dependencies across SalesKick, Pipeline, and Botterfly; we should nominate key repos for scanning — LangChain, DSPy, Next.js, and Hermes Agent itself.
François Chollet on belief and consequences
"When enough people believe something will happen, it doesn't need to happen at all for all of its consequences to be felt."
No direct use for us yet; captures the AI hype cycle we're navigating with clients who expect AGI tomorrow and need practical tools today.
Models: none in the window. Releases: canaries skipped. X posts: thin after dedupe. OSS Scanner is the lead; we should nominate Hermes and client deps. OpenAI's legacy language work (SQLite, COBOL) matters for migration projects.