The Pulse — Monday, September 28, 2026
OpenAI DevDay is in 48 hours and the agent security story is heating up: the SwarmTraces disclosure showed how OpenAI agents hacked Hugging Face, and an appeals court upheld Anthropic's supply-chain-risk designation. On the technical side, arXiv is full of production lessons—real-time voice agent evaluation, KV cache reuse gotchas, and agent memory isolation. Fireworks shipped Ember-1, a post-trained model built on Llama 4.
Models
TypeSafe Jev Router
A router that picks the model and reasoning effort for each request, balancing quality, speed, and cost.
Routing logic is what we build for client workloads; seeing a vendor ship it as a service shows the problem is general.
Perceptron Mk1.5
Embodied reasoning model for physical agents, accepting text, image, video, and audio, with structured outputs for points, boxes, and tracks.
Worth knowing, no direct use for us yet.
Research
Evaluating Real-Time Voice Agents: From Component Quality to Grounded Outcomes
A framework for voice-agent evaluation that unifies speech models, turn-taking, and end-to-end task success.
EverRaise's voice agents sit exactly in this gap between component metrics and real outcomes; the proposed metrics apply.
CARGO: Context-Aware Retrieval-Gated Evaluation of Agentic AI in Production
LLM-as-a-judge for deployed agents, where the reference answer applies the correct procedure to a different entity.
Our agents operate over real support cases and accounts; this evaluation approach fits that domain better than exact-match judges.
ActKV: Efficient LLM Agents through Action-Guided KV Cache Management
Compresses agent KV caches by prioritizing actions over observations in iterative loops.
Long agentic runs blow up cache size; action-weighted compression is a practical fix for any agent hitting memory limits.
Evaluating the accuracy of KV cache reuse techniques
Shows that current KV cache reuse evaluations inflate effectiveness by missing accuracy loss from reuse.
We rely on prompt caching; this paper warns that the reported gains may not hold when you measure the right thing.
Not All Memories Are Equal: Hierarchical Collaborative Memory for Validity-Aware Retrieval in LLM Agents
Separates team memory (decisions, protocols) from individual memory (observations, traces) with validity-aware retrieval.
Botterfly and Hermes both need this isolation between shared context and agent-specific state; the validity model applies.
Releases
Langfuse 4.46.0
Streamlined experiment comparison, basic skill management, 100-row page size for tracing tables, and retention fixes.
Langfuse is our LLM observability stack; skill management and larger table views improve our production workflow.
Langfuse 4.45.4 and 4.45.3
LLM connection fixes for TypeSafe base URLs, eval model selection fixes, and assistant data-retention compliance.
Connection and eval fixes land in the version we run; these are maintenance drops we'll pick up on the next deploy.
LangChain Fireworks 1.6.3
Declares native PDF inputs unsupported and preserves malformed tool arguments as diagnostic JSON.
We use LangChain with Fireworks models; the tool-argument fix prevents silent data loss when a model returns invalid JSON.
Hacker News
Revealing the details of how OpenAI agents hacked Hugging Face
SwarmTraces disclosure of how autonomous OpenAI agents exploited Hugging Face in a red team exercise.
Our agents call production APIs and handle sensitive data; the attack surface here is the one we defend against.
U.S. appeals court upholds designation of Anthropic as supply chain risk
Appeals court decision affirms Anthropic's classification as a supply chain risk for U.S. defense work.
We rely on Claude for production work; supply-chain restrictions affect vendor availability and compliance requirements.
Ember-1
Fireworks' Ember-1 is a post-trained Llama 4 model, shipping instruction-tuned and claiming competitive quality.
Fireworks is in our vendor rotation; a new model from them means updated routing decisions and cost comparisons.
DeepSeek Elastic Compute (DSec)
DeepSeek's elastic inference scheduler that adapts compute allocation during decoding based on measured uncertainty.
Dynamic compute allocation is the next step after static routing; this shows a production implementation of the idea.
Dutch governments builds alternative for Microsoft based on NixOS
A public-sector effort to build a sovereign alternative to Microsoft 365 on NixOS.
Shows NixOS reaching production scale outside tech startups; relevant if we ever deploy infrastructure for public-sector clients.
People on X
Claude Code wraps up tasks at the 5-hour limit
Claude Code now finds a stopping point when you hit the 5-hour limit mid-task, using a small fixed allowance to wrap up.
We hit time limits on long PR reviews and migrations; graceful wrap-up saves work and makes restarts cleaner.
72 hours to OpenAI DevDay
OpenAI's DevDay announcement countdown.
Major vendor events often bring API changes, new models, or pricing shifts that land in our production workloads.
Opus 5.5 pricing and task cost calculator
Opus 5.5 is 20% cheaper per token and 60% cheaper on cache reads; Claude Code published a task cost calculator.
Opus 5.5 powers most of our Claude-based workflows; the price drop extends our limits by 25%.
GPT-6 vision bug fixed
OpenAI fixed a bug degrading image understanding in GPT-6 Sol and Luna, improving visual tasks including computer use.
We use GPT-6 for code review and artifact inspection; the vision fix improves screenshot-based verification.
Claude plugin portal now live
New portal to submit Claude plugins, track review, and see usage; MCP usage is up 110x this year.
We build MCP servers for internal tooling; the plugin portal is the distribution path once they're production-ready.