Friday, July 17, 2026

Claude Code Rewrites Migration Playbook

Claude Code Rewrites Migration Playbook

Today’s Overview

Good morning, Claude Code is turning code migrations into a more mechanical, agent-driven process, while Moonshot is pushing open models into 3T-class territory with Kimi K3. There’s also a wave of agent infrastructure, from LM Studio’s Bionic app to Perplexity’s secure sandbox work. Let’s dive in.

Top Stories

Moonshot Launches Kimi K3

Moonshot has launched Kimi K3, a 2.8-trillion-parameter multimodal model with a 1-million-token context window. It is built for long-context work, agentic coding, and multimodal reasoning, and is available through Kimi products and the Kimi API. Open weights are scheduled for release on July 27.

  • Architecture-wise, Kimi K3 uses Kimi Delta Attention and Attention Residuals to improve information flow across long sequences and model depth.
  • Its sparse MoE setup activates 16 of 896 experts using Stable LatentMoE, with Moonshot claiming roughly 2.5x better scaling efficiency than Kimi K2.
  • Moonshot says the API pricing starts at $0.30 per million cache-hit input tokens with cache-miss input at $3.00 and output at $15.00 per million tokens.

Anthropic Turns Code Migration Into an Agent Workflow

Anthropic outlines a six-step process for using Claude Code on large-scale code migrations. The workflow centers on a migration rulebook, dependency analysis, adversarial review, and mechanical checks that compare the new code against the original. The approach uses multiple agents to translate, review, and fix code iteratively rather than treating migration as a one-shot rewrite.

  • One example migrated 165,000 lines of Python to TypeScript over a weekend using hundreds of agents, eight phase gates, and three adversarial review rounds.
  • The process starts by building a judge that can test both codebases equally so the team has a concrete exit condition instead of relying on subjective review.
  • Anthropic recommends reserving larger models for reviewers and rule-writing while smaller models handle high-volume implementation fan-out.

LM Studio Introduces Bionic for Open-Model Agents

LM Studio introduced Bionic, an AI agent for coding, research, and document work built around open models. It supports local and cloud execution so users can balance privacy, model capability, and cost. The app also includes offline voice transcription, document workflows, sandboxed processing, and native web search.

  • LM Studio says Bionic users get Zero Data Retention and that user data is not used for training.
  • For voice input, Bionic ships with Voxtral by Mistral AI for local multilingual real-time transcription.
  • For coding workflows, Bionic can inspect local codebases then investigate, edit, debug, and show inline diffs for review.

Research & Analysis

SearchOS Makes Agent Search State Persistent

SearchOS reframes open-domain information seeking as relational schema completion with grounded citations. Its Search-Oriented Context Management approach externalizes agent state into structures such as a frontier task list, evidence graph, coverage map, and failure memory. The system is designed to reduce repetitive search loops and improve collaboration across long-horizon web-search agents.

  • The framework treats answers as linked table completion where discovered entities and attributes are anchored to source evidence.
  • Its middleware harness records model and tool interactions so the system can react to stalls, missing evidence, or budget exhaustion.
  • The paper reports SearchOS leading all evaluated metrics on WideSearch and GISA among the tested single-agent and multi-agent baselines.

LongStraw Pushes Long-Context RL Past 2M Tokens

LongStraw is an architecture-aware execution stack for million-token reinforcement learning post-training under a fixed GPU budget. It is instantiated with Group Relative Policy Optimization and designed for agentic workloads where observations, tool outputs, and previous decisions accumulate over very long contexts. The approach reduces the live training graph by replaying short response branches one at a time, trading memory pressure for extra replay time.

  • On eight H20 GPUs, LongStraw completes grouped Qwen scoring and response backward at 2.1M positions for groups of 2 and 8.
  • Increasing group size adds only 0.21 GB peak memory in the reported Qwen experiments.
  • A separate stress test reaches 4.46M positions showing the execution stack can stretch beyond the main 2.1M-token setup.

SEED Turns Agent Trajectories Into Skills

SEED is a self-evolving reinforcement learning framework for agentic language models. It converts completed on-policy trajectories into natural-language hindsight skills, then distills their behavioral effect back into the policy. The goal is to bridge sparse episode-level rewards and dense token-level learning for more sample-efficient agent training.

  • The policy model is trained to analyze completed trajectories and extract reusable workflows, decisive observations, or failure-avoidance rules.
  • During RL, the same policy both collects trajectories and acts as the analyzer that generates hindsight skills from those trajectories.
  • SEED converts skill effects into dense token-level signals by comparing action probabilities under ordinary and skill-augmented contexts.

VideoChat3 Targets Efficient Open Video Understanding

VideoChat3 is a fully open video-centric multimodal language model for general-purpose video understanding. It combines an Inflated 3D Vision Transformer backbone with adaptive frame resolution for streaming video perception. The paper positions the 4B-parameter model as a balance of broad video generalization and compute efficiency.

  • The training pipeline curates three datasets covering general, long-form, and streaming video scenarios.
  • Those datasets are named VideoChat3-Academic2M, VideoChat3-LV116K, and VideoChat3-OL617K and are intended to improve cross-domain generalization.
  • The Hugging Face page lists both VideoChat3-4B and I3D-ViT as models associated with the paper.

Trending AI Tools

  • dd-cli DoorDash’s beta AI agent tool for finding deals, searching restaurants, and placing orders.

  • Copilot SDK A GitHub SDK for embedding Copilot agents into custom applications and developer tools.

  • SPACE Perplexity’s sandbox platform for agents handling sensitive tasks with ephemeral execution and credential isolation.

Quick Hits

  • Apple Intelligence in China has reportedly been registered and approved for iPhones, while Siri still needs additional regulatory steps before launch.

  • Gemini Notebook is the new name for NotebookLM, with deeper ties into the Gemini app and Google Search.

  • Basedash Suggestions positions the product as an AI data analyst that can proactively generate ideas, not just answer queries.

  • Connected Apps in AI Mode adds Instacart, Canva, and YouTube integrations so users can complete tasks across supported apps from conversational search.

Keep reading for free

Enter your email. If you're already subscribed, we'll send a sign-in code. If not, you'll subscribe in the next step.

Free access. Subscribe once, then use the same email on future issues.

Free to read. Subscription just unlocks the full issue.