The Pulse, Monday 19 October 2026
DeepSeek shipped V4.1 Flash 1017 with vision and 1M context for 15¢/60¢ per million tokens. Claude Code's team wrote about how they do code review when the agent writes 90% of the code. The weekend brought three days' worth of agentic memory papers, CLI benchmarks, and safety research.
Models
DeepSeek: DeepSeek V4.1 Flash 1017
DeepSeek V4.1 Flash 1017 is the October 17 snapshot of DeepSeek's sparse mixture-of-experts model, now with vision (text+image->text), 1M context, and improved reasoning.
Matches Claude 3.7 performance on SWE-bench Verified at a tenth of the cost; the whole Pipeline stack runs through OpenRouter so this is immediately available for all our agentic call grading and retrieval workflows.
Elorian: Glimpse
Glimpse is a visual reasoning model from Elorian built for physical-world understanding; it handles images and video (text+image+video->text) with 262K context.
Video understanding for EverRaise's call recordings and Life International's training materials; the multimodal reasoning could help Meno's bill-tracking agents verify document changes across video briefings.
Reflection: Beam 1.0 501B A23B
Beam 1.0 is Reflection's open-weight reasoning model for coding and long-horizon agentic workflows, 501B total with 23B active per token and 1M context.
Built specifically for agentic coding workflows; worth testing on Botterfly's skill authoring and wehshi review panel tasks where reasoning depth matters more than raw speed.
NVIDIA Nemotron Labs Dory1.0 4B
Dory1.0 is a 4B conversational model from NVIDIA's Nemotron Labs, released in BF16, MXFP8, NVFP4, and GGUF quantizations.
Small enough to run on-device for SalesKick's call grading pilots on RTX laptops; the GGUF release means we can eval it in llama.cpp before committing to fine-tuning.
Research
AuditKIT: A Unified Evaluation Spine for Benchmarking, Judging, and Comparing Large Language Models and Agentic Systems
AuditKIT combines benchmark accuracy, judge-based assessment, runtime performance, tool-use behavior, and retrieval quality into one evaluation framework for LLM agents.
Exactly what we need for systematic comparison of wehshi review models and EverRaise's Vapi-based voice agents; the unified trace format could feed our Langfuse observability stack.
MetaCurate: Evolving Curation Filters for High-Quality Agentic Coding Data
MetaCurate learns to filter agentic coding trajectories by distinguishing useful intermediate steps from noise, even in failed runs.
Our wehshi panel and Claude Code sessions generate thousands of trajectories; this could help us curate the keepers for fine-tuning Botterfly's skill authoring and code review behavior.
SAFE-Mem: Secure Agentic Memory with Fine-Grained Authority and Governance-Aware Retrieval
SAFE-Mem prevents untrusted claims in LLM agent memory from being treated as authorized preferences or permissions.
Hermes Agent's Mnemosyne memory lets agents store facts from external sources; this addresses the exact security gap we saw when a Pipeline client's scraped data was mistaken for a preference.
Ignored, Not Refused: What Low Prompt-Injection Success Rates Hide in Small Tool-Using Agents
Audit of 33K trajectories shows that prompt-injection defenses work by ignoring the attack, but the agent still exposes the injected content through other tool calls.
Every Pipeline voice agent reads untrusted caller input and calls tools; if our defense is silent ignoring, the attack still leaks through context — this tells us what to test.
From Relevant Memories to Sufficient Evidence: Rethinking Memory Utilization in LLM Agents
Retrieving relevant memories doesn't guarantee sufficient evidence for reasoning; essential information stays fragmented across memory records, obscured by abstraction.
Mnemosyne's hybrid recall pulls memories by similarity, but this shows why our agents still miss facts — relevance ≠ sufficiency; the paper's evidence-chaining method could improve Botterfly's decision recall.
Releases
LangChain Core 1.6.12
Fixes extra propagation on content-block-start events.
We use LangChain Core across SalesKick's nurturing agents; the content-block-start fix unblocks streaming for Mackinac's research workflows where we show partial results.
Langfuse 4.63.0
Ships skill management as a public v4 preview, AI gateway codex-model routing, and annotation queue UX improvements.
We've been waiting for public skill management to systematize Botterfly's procedural memory; the codex-model routing lets us send slugless models to the OpenAI connection without manual mapping.
Langfuse 4.62.0
Adds AI gateway per-connection model controls, request limit raised to 16 MiB, and OTEL third-party span metadata decoding.
The 16 MiB request limit unblocks EverRaise's longer call transcripts through the gateway; per-connection model controls let us route client-specific provider credentials cleanly.
Langfuse 4.61.0
Adds dedicated skills tab on traces, AI gateway connection-based routing, and skill usage metrics.
Skills-on-traces visibility means we can finally see which Botterfly skills fired for each wehshi panel run; the metrics will show which skills are dead weight.
Hacker News
Claude reveals ECDSA curve origins
Anthropic used Claude to trace the provenance of the ECDSA P-256 curve constants back to deliberate parameter choices in 1991 NIST research, not unexplained magic numbers.
Shows a reasoning model can reconstruct cryptographic design decisions from decades-old papers; the same investigative approach could help us audit black-box third-party SDKs in client stacks.
How the Claude Code team does code review
Claude Code writes 90% of its own code; the team shifted code review from line-by-line diffs to automated verification, self-fixing agents, and human focus on design decisions only.
Exactly the problem we're solving with wehshi: when agents write faster than we can review, verification becomes the bottleneck — their playbook (gate on tests, autofix lints, verify claims) directly informs our PR review pipeline.
DeepSeek search and fetch – $0.75 per 1000 successful requests
DeepSeek launched a search and fetch API at 75¢ per 1K successful requests; it returns search results as structured JSON for LLM consumption.
Mackinac's legislative research and EverRaise's donor intelligence both need live web search; at 75¢ per thousand this undercuts our current Serper and Brave bills by 80% — test immediately.
OpenCode Hacked: User credentials leaked
OpenCode disclosed a breach that leaked user credentials; the service was temporarily shut down while the team rotated keys and notified users.
OpenCode is referenced in our skills and deployment docs; we need to audit any OpenCode integrations across SalesKick and Pipeline for credential exposure and rotate affected keys.
People on X
@ClaudeDevs on performance improvements
Claude Code now gzips the conversation before sending it to the model on every turn, making the agent start responding sooner; cloud sessions also start as soon as typing begins.
Latency matters for interactive agents; we can borrow the gzip-before-send trick for Hermes sessions with long Mnemosyne recall or sprawling skill contexts.
@fchollet on software engineering in 2026
François Chollet observes that there are more software engineers now than in 2025, demand is growing, they work harder than ever, and agents write ~100% of the code.
Matches what we're seeing: Botterfly and wehshi write the code, but we're busier because the bottleneck moved to design, verification, and integrating what the agents built — no direct use for us yet.
@NousResearch on Hermes Agent and NVIDIA RTX Spark
Hermes Agent is a launch partner for NVIDIA's RTX Spark across the new generation of laptops; users can pick it in the built-in Windows agent setup flow.
Hermes is our daily driver for Botterfly and all Pipeline work; RTX Spark partnership means clients buying new dev machines get Hermes out of the box — good for adoption, and Linus is impressed.
Python releases: nothing notable in the window. DeepSeek and Claude Code dominated the weekend; expect a lot of teams testing V4.1 Flash this week.