Can You Prove Which Model Answered?

Claude Code 2.1.323 adds an experimental Code Mode that writes JavaScript to call its tools, plus a batch of permission fixes worth updating for. Thinking Machines releases open weights for Inkling 2 Small, llama.cpp and ggml hit 1.0 with a semver promise, IBM's Granite LLMON fights prompt injection with tagged prompt blocks, and Anthropic opens a private beta of self-hosted Managed Agents. We teach why fewer tool-call round trips matter and how mixture-of-experts models can be huge yet cheap to run.Hosts: Alex & Jules. New episodes daily.

AWS releases AWS Emulator, a free offline copy of 200+ services written entirely by agents and checked against real AWS, while Google's Gemini Decisions API and Amazon's open-source Strands Decider turn yes/no and multiple-choice calls into millisecond, near-free decisions. We teach why scoring options in a single forward pass is so much cheaper than generating an answer token by token. Plus Chrome's Modern Web Guidance skills for coding agents, Qwen4-Max-Preview's Anthropic-compatible API, Claude Code 2.1.322, Apple's rumored Harbor app, and United's ban on booking agents. Hosts: Alex & Jules. New episodes daily.

Gemini 4 Carbon opens to paid developers at introductory pricing, and we teach how multi-agent orchestration works through Antigravity's new /swarm feature. Railway's post-mortem puts stolen AI key usage at about $740,000, and we explain envelope encryption and why it didn't stop the attack. Plus Adobe's Creative Cloud MCP server, ElevenLabs Scribe v3 Live with realtime diarization, a one-time Claude usage reset, and Qwen4-Max in preview.Hosts: Alex & Jules. New episodes daily.

Anthropic's Claude Fable 5.2 arrives at a lower price and becomes Claude Code's default Fable, alongside new security fixes in 2.1.321. We teach why per-token price misleads and how input, thinking tokens and cache hits set the real cost per task, then explain confidential computing via Google's TPU 8tx and its open-source TPU agent suite for Claude Code and Cursor. Plus Anthropic's connector-powered office suite and Qwen3.8 topping Hugging Face.Hosts: Alex & Jules. New episodes daily.

Claude Code 2.1.319 adds background MCP tasks and a deny-instead-of-prompting setting for headless runs, Cursor adds realtime voice chat with agents, and AWS ships persistent Kiro bots. We also cover Meta's Conductor acquisition and DigitalOcean's Cuttlefish iPhone app, and teach the difference between blocking and background tool calls and between agent memory and model retraining. Hosts: Alex & Jules. New episodes daily.

Apiiro reports thousands of secrets leaking from Claude Code's "Don't Ask Again" approvals file, and we cover what to rotate. Claude Code 2.1.319 adds background MCP tasks and deny-instead-of-prompt, and we teach how agent memory works, including Anthropic's "dreaming" research. Plus NVIDIA's Vera CPU for agents. Hosts: Alex & Jules. New episodes daily.

A LangChain study claims Stagehand beat Playwright MCP on cost, and we teach how caching lets an agent skip repeat model calls. Plus Phone Harness gives Claude Code a real phone, Stacktrace and CrowdStrike ship agent controls, and a Claude Code fast-mode deadline on November 30. Hosts: Alex & Jules. New episodes daily.

Cursor Apps turns a description into a hosted app with database and auth, Antigravity CLI plugs into any editor via ACP, and Codex CLI, Copilot CLI, Amp and Claude Code all ship updates. We also teach how the Agent Client Protocol (ACP) lets editors and coding agents talk over JSON-RPC.Hosts: Alex & Jules. New episodes daily.

Claude Code 2.1.317 adds a prompt before your account email goes out over the network, and we explain why anything in a model's context can leak into its outputs. We also cover Stroq, an open source action firewall for coding agents (and the idea of "tainted" sessions), Tencent's Hy4 model, and how hybrid attention layers cut long-context memory costs.Hosts: Alex & Jules. New episodes daily.

Anthropic's FlashProve prototype aims to cryptographically prove which model produced a response, Moonshot releases Kimi K3.1 Pro weights (2.8T parameters, 104B active), and we cover MCP OAuth redirect risks, Gemini's chat-training settings, and an RTX Spark Linux slowdown. Today we teach mixture-of-experts and zero-knowledge proofs in plain terms. Hosts: Alex & Jules. New episodes daily.