The Pulse, Monday 12 October 2026
Cloudflare acquires Deno, a direct signal for our EverRaise edge work. Claude rolled out dynamic workflows for Managed Agents and opened Code Projects to all Pro/Max users. The weekend brought 40 arXiv papers including a sales-agent benchmark that tests objection handling and a Friday research haul on coding agents.
Models
Qwen-Image-2.1-Turbo
Qwen released an accelerated image generation checkpoint that creates and edits images in 8 denoising steps, open weights on HuggingFace.
Worth knowing, no direct use for us yet.
Claude Managed Agents dynamic workflows
Anthropic shipped dynamic workflows in beta: a lead agent writes a plan that runs across many agents in phases and combines results at the end.
Multi-agent orchestration pattern for large tasks. Relevant to Botterfly's own review panels and any SalesKick workflows that need parallel evidence gathering.
Research
SmartSalesBench: A Resistance-Aware Multi-Turn Benchmark for Evaluating LLM Sales Agents
First multi-turn conversational benchmark designed to evaluate sales agents against psychologically grounded, resistant buyer personas instead of cooperative ones.
Directly relevant to SalesKick. Objection handling and non-cooperative conversations are the hard part of sales coaching; we can run this benchmark against our own agent designs.
MADS: Multi-Agent Dialog Simulation for Robust Conversational Agents
Framework for simulating adversarial conversation trajectories to train robust task-oriented conversational agents under non-cooperative user behavior.
Training data generation for EverRaise voice agents and SalesKick call simulations. Our Vapi agents need to handle off-script user behavior; this gives us a method to generate those test cases.
Rollout-Time Intervention: Unlocking Agentic Reinforcement Learning Beyond Policy Limits
RL method that lets agents improve reasoning and tool-use during rollout by intervening when tasks exceed the model's current capability boundary.
Potential improvement path for agents that get stuck. Relevant to Botterfly's autonomous coding workflows and any long-running SalesKick coaching sessions.
SWE-Duel: Towards a Self-Scaling Arena for Coding Agents
Self-scaling benchmark where tasks and oracles are created at evaluation time so content never leaks into training data, and difficulty adapts to participating agents.
Prevents benchmark saturation for Botterfly and our coding eval pipeline. If we're evaluating agent improvements, this gives us a benchmark that won't go stale.
Releases
Python 3.15.0
Python 3.15.0 stable released.
Our entire backend stack. SalesKick services, Pipeline's ai-service, Botterfly scripts, and every agent harness we run.
Langfuse v4.56.0
Langfuse v4.56.0 adds syntax highlighting to session tool previews, adaptive map exploration for topics, revamped eval decision model results, and spend alerts for ClickHouse-billed orgs.
Observability for SalesKick and EverRaise agents. We rely on Langfuse for tracing multi-turn voice and coaching sessions.
Hacker News
Cloudflare acquires Deno
Cloudflare acquired Deno, the TypeScript-first runtime and edge platform.
EverRaise edge functions run on Deno Deploy. This consolidates our edge stack under one vendor and signals continued investment in the Deno runtime we're already committed to.
Telegram Desktop vulnerability allowed any user's file to be stolen
Telegram Desktop had a one-click account takeover vulnerability that allowed arbitrary file exfiltration from any user's machine.
Reminder to audit file-access patterns in any client app. Relevant to tooling and integrations we build with filesystem access.
Bitwarden Dual License Model
Bitwarden announced a dual license model in its latest update, moving portions of the codebase away from GPL.
License risk for any self-hosted password management we consider. Not blocking but worth tracking if we ever recommend or integrate Bitwarden for client projects.
OpenAI fires three safety researchers for "mishandling research information"
OpenAI fired three safety researchers for alleged misconduct; the researchers dispute the claims and warn of a chilling effect on safety work.
Industry signal on safety team dynamics at a vendor we rely on for SalesKick and EverRaise. No immediate impact but relevant to vendor risk.
People on X
Claude Code Projects now available to all Pro and Max users
Claude Code Projects is now available to all Pro and Max users, with a 4-minute walkthrough for getting started.
The agent tooling we're already using for Botterfly and client work. Projects mode keeps related tasks coordinated across sessions.
shadcn on reading as the most important coding skill
The most important coding skill: read. Read the agent output, the blog post, the paper, the code. Don't summarize.
Craft reminder for our own agent workflows. When Botterfly or any of us review agent output, reading the raw result beats asking for a summary.
François Chollet on AI paradigm shift
AI shifted from transductive (token-by-token completion) to inductive (on-the-fly reasoning chain synthesis). This paradigm shift is what made coding agents work.
Explains why agents like Claude Code and our own Botterfly work now when they didn't a year ago. Reasoning chains synthesized at inference time are the unlock.
Nous Research teases MISSINGNO model
Nous Research announced MISSINGNO, a coding and agentic reasoning model that is highly token efficient, free for limited time on Nous Portal.
New coding model from Nous, relevant to our Hermes Agent work and worth testing against Botterfly's own tasks.
Sebastian Raschka on RLVR (Reinforcement Learning with Verifiable Rewards)
Sebastian Raschka's second RLVR tutorial covers clipped policy ratios, KL loss term, format rewards, and other GRPO tips and tricks.
Training technique for reasoning models. Relevant if we experiment with post-training for SalesKick coaching models or Botterfly's own reasoning improvements.
Next.js had only canaries in the window; nothing notable to report. OpenAI vendor feed announced Sophos and Asana case studies but no model or API changes relevant to our work.