Intel
Page 05
Claude Code Code Mode: Tool Calls as Scripts
Claude Code gains an experimental Code Mode: Claude writes JavaScript that runs tool calls in parallel and feeds results forward, optionally in the background.
Why it mattersCode Mode lets Claude Code write one JavaScript script that chains, parallelizes and filters tool calls instead of reading each result in turn. Docs list real limits: auto mode reviews the code, not each call, and hooks behave differently.

Harvard researchers find AI coding agents shifted the bottleneck to code review
A Harvard working paper across 718 firms finds AI coding agents added about 30% more code while finished work did not measurably rise and review took 49% longer.
Why it mattersCoding agents raised code volume without raising finished work, because review became the bottleneck. Teams adopting agents should measure review time and approved output, not just lines drafted.
Artificial Analysis reports GPT-6.2 Astra at number one on AA-3D, its benchmark for building interactive 3D experiences, with humans preferring its scenes in 139 of 152 head-to-head votes. Leaderboard details and the demo video were not available to verify.
Why it mattersIf you generate interactive 3D scenes with models, this gives a human-preference signal: GPT-6.2 Astra won 139 of 152 pairwise votes on AA-3D. Treat it as a single benchmark with a small vote count.
At Ignite, Microsoft announced Copilot Workspaces for shared human-agent context, Decision 1.1 GA, Code 2, Cyber 2 Flash, GPT-6.2 and Meta models in Foundry, Agent 365 analytics, and Perception GA. Performance claims are unverified.
Why it mattersGPT-6.2 and Meta's latest models are coming to Foundry, and Microsoft shipped its own coding and security models plus shared-context agent workspaces. This changes which models and agent tooling are available on Azure.
A Background Agent to Fix Breaking Changes in Dependency Updates
A walkthrough of a GitHub Action that runs a coding agent when a dependency-update PR breaks tests.
Why it mattersShows how to bound a CI repair agent with budget, timeout and ignore-scripts so Renovate PRs with breaking changes get fixes pushed for review instead of piling up.
Perplexity launches the Orchestrator API: 45+ models from OpenAI, Anthropic, Google, xAI and open-weight providers behind one key, callable via OpenAI and Anthropic SDK formats. Preview auto-routing picks a model per request, with per-token billing, spend caps and allowlists. The video was not trans
Why it mattersOne API key can reach 45+ proprietary and open-weight models through existing OpenAI or Anthropic SDKs, with per-request model routing and spend controls.
Mollick had Astra 6.2 build a Blender and Unity time-travel game. The result was impressive and playable, but the model added unrequested mechanics, skipped a shooting mechanism, and needed constant redirection, unlike GPT-6 Asta.
Why it mattersDescribes a specific steerability failure: Astra 6.2 builds unrequested features without telling you, which raises supervision cost in agentic coding even when output quality is high.

2026 in LLMs — Simon Willison, Datasette
Simon Willison's CODE Summit 2026 talk traces three inflection points in LLMs: reliable coding agents, persistent problem solving, and vibe-coded games.
Why it mattersA practitioner's view of when coding agents became reliable, how they handle persistent problem solving, and the security risks of prompt injection and rogue agents.
Scribe V3 Live Medical
ElevenLabs releases Scribe v3 Live Medical, a real-time speech-to-text model for healthcare voice agents in ElevenAgents.
Why it mattersVoice agents in healthcare can now use a real-time transcription model tuned for drug names and clinical terms, with up to 10,000 custom terms, reducing misheard medications during live calls.

New Ways of Working with Codex — Dominik Kundel & Charlie Guo, OpenAI
OpenAI's Dominik Kundel and Charlie Guo demo connected Codex workflows at CODE Summit 2026: plugins, Appshots, Codex Cloud, ChatGPT Sites, voice, code review, and parallel agents.
Why it mattersSee how OpenAI's own team connects Codex to plugins, computer use, cloud environments, code review, and parallel agents, including model and reasoning-effort choices.
OpenRouter adds access controls to management keys: make a key read-only, limit it to one workspace, or both. Available on all plans, so each agent, script or tool can hold only the access it needs and a leaked key has a smaller blast radius. The linked docs were not read.
Why it mattersOpenRouter management keys can now be restricted to read-only and to a single workspace on all plans. Agents and scripts can run with least privilege, so a leaked key exposes less of the account.
Oh My Pi, an open-source terminal coding agent, now lets its judge role run on Nace's Drex decision models. The judge returns calibrated probabilities for typed questions, and Drex Max can read screenshots. Setup needs omp 18.20.0 or later and a config change.
Why it mattersOh My Pi users can now assign a dedicated decision model to the judge role, getting calibrated probabilities for small typed questions while the main chat model stays unchanged. Drex Max also reads screenshots.
Google DeepMind announces a robotics plugin for Antigravity paired with Gemini 4 Hydrogen. Users connect a supported arm, practice in MuJoCo simulation, and run tasks on real hardware through plain conversation. The post does not list supported arms or give performance data.
Why it mattersA coding-agent environment now has a robotics plugin: you can connect a supported arm, rehearse tasks in MuJoCo simulation, then run them on real hardware through conversation. Supported arms and limits are not specified in the post.

llm 0.38
Simon Willison's llm CLI adds OpenAI's new gpt-6.2-astra and gpt-6.2-sol models, speeds up non-OpenAI commands via lazy imports.
Why it mattersIf you use the llm CLI to call GPT-6.2 Astra/Sol or run async OpenAI prompts with response logging, this release removes a crash and restores faster startup for unrelated commands.

D.a.r.y.l. (1985), a film I loved as a kid, as an AI system card
A blog post recasts the 1985 film D.A.R.Y.L. as an AI system card, pairing each scene with real model results on sandbagging, shutdown resistance and escape attempts.
Why it mattersCollects cited evaluation results on sandbagging, shutdown resistance and exfiltration across several models in one place, useful for engineers reasoning about agent failure modes.
Epoch AI's pre-launch evaluation of GPT-6.2 Astra: ECI 174 on par with Fable 5.2, weaker on realistic research tasks, 12.3 serial reasoning steps without chain of thought, and a record 7/68 Lean-verified FrontierMath Erdős solutions at about $240 per question.
Why it mattersEpoch's own benchmarks show GPT-6.2 Astra tying Fable 5.2 on ECI but trailing on realistic research tasks. It also does more serial reasoning without chain of thought (12.3 steps vs 4.2 for Fable 5.1), which may make it harder to monitor.

openai-python v3.53.0
openai-python 3.53.0 adds model identifiers for GPT-6.2 Sol, Astra, and Astra Law, plus 2-hour and 8-hour background windows for the Responses API.
Why it mattersThe OpenAI Python SDK now exposes GPT-6.2 Sol, Astra, and Astra Law model IDs and longer 2-hour and 8-hour background windows for Responses, so long-running agent jobs can run in background mode.

ai ai@7.0.174
Vercel AI SDK 7.0.174 adds output support in startBatch and getBatchResults, extraction model support in provider registries, and fixes streamText transform factories.
Why it mattersIf you build on the Vercel AI SDK, this release adds batch output parsing and extraction model registry support. It also fixes a validation path where deeply nested metadata could exhaust worker memory.

ai ai@6.0.333
AI SDK 6.0.333 reduces memory and CPU use when validating UI messages with deeply nested provider or tool metadata, validating in a single pass without copying values except to strip __proto__ keys.
Why it mattersIf you run AI SDK 6 servers, this patch removes a validation path where a crafted nested-metadata request could consume excess memory and CPU before the model is even called.
Theo's full letter to OpenAI on GPT-6.2 Astra: strong reviewing, computer use and 3D work, but self-blocking questions, unreadable explanations and loops. It landed no improvements in 11 days on a ts-rust task where Opus 5.5 finished.
Why it mattersDocuments concrete long-run failure modes of Astra 6.2: self-blocking on internal questions, unreadable summaries, and 11 days of zero landed improvements on a compiler task. Useful when deciding which model drives long agent runs.

Maintainers don’t need more AI bug reports. They need fixes.
GitHub and Microsoft piloted Codename MDASH, an agentic security scanner.
Why it mattersLays out a concrete find-filter-fix pipeline for responsibly using AI security scanners against open source, directly relevant if you're deploying or receiving AI-generated vulnerability reports.

anthropic-sdk-typescript sdk-v1.10.0
Anthropic TypeScript SDK 1.10.0 adds beta billing_policy support, contract spend limit and linked-billing error codes, and webhook events for external key failure and recovery.
Why it mattersAgent builders on the Anthropic TypeScript SDK get a beta billing_policy parameter, new billing error codes, and key failure webhooks. Session tool runners also stop executing tools for terminated sessions.

MAI Code 2 and Code 2 Flash
Microsoft AI released MAI Code 2 for multi-file implementation, debugging and repo-wide reasoning.
Why it mattersMicrosoft AI added two coding models, one for multi-file and repo-wide work and a cheaper Flash tier for high-volume edits, giving engineers another cost option for coding agents. Verify benchmark and price details at the source.

Building the Agent Host Protocol for sessions that continue across devices
GitHub details the Agent Host Protocol, an open standard that decouples a long-running agent session's lifetime from any single editor window.
Why it mattersA concrete architecture pattern for building agent sessions that survive editor restarts and follow you across devices, directly applicable if you're building your own multi-client agent product.
Ethan Mollick reports GPT-6.2 Astra was uneven in pre-launch use: it delivered a playable space elevator world built with Blender and Unity, but required more steering than competing models and felt less like a clear step forward.
Why it mattersEarly hands-on evidence that Astra 6.2 can deliver ambitious builds but needs more steering than other models, relevant when choosing a model for agentic work.
Every's early hands-on report on OpenAI's Astra 6.2 and Sol 6.2 after weeks of testing: coding similar to Astra 6 on benchmarks but stops early in real builds, terse writing, a recognizable design house style. Preliminary; the full review is pending.
Why it mattersIf you build end-to-end with coding agents, a reported tendency of Astra 6.2 to stop early on full builds is worth checking before you switch from Astra 6. The claims are an early impression from one team, not benchmarked results.
Artificial Analysis benchmarks Microsoft AI's MAI-Image-2.7 and Flash: #1 in image editing, #4 in text-to-image, priced at $34.82 and $10.24 per 1k images. Both are in public preview and move the quality-versus-price frontier.
Why it mattersMAI-Image-2.7 ranks #1 on AA-Image-Editing v2.1 at $34.82 per 1k images, with Flash at $10.24. Teams building image-editing pipelines get a cheaper option, though results come from one benchmarker's square-output tests.
OpenRouter announces GPT-6.2 Astra and GPT-6.2 Sol from OpenAI are live, with Pro variants of both. Astra is positioned for the hardest projects and Sol for everyday agentic work. The post gives no pricing, context limits or benchmark data.
Why it mattersGPT-6.2 Astra, Sol and their Pro variants can now be called through OpenRouter. Teams routing across providers can test them against their current models, but the post gives no pricing or limits.
OpenAIDevs says GPT-6.2 Astra and Sol are now usable in Codex and the API. A demo transcript describes Astra recreating a filmed room in Blender from two frames and iterating on it. The demo is promotional, visuals were not analyzed, and no pricing or benchmarks were given.
Why it mattersGPT-6.2 Astra and Sol can now be called in Codex and the API. OpenAI's demo claims Astra rebuilt a room in Blender from two video frames, but no benchmarks or pricing were published in this post.
Vals AI's pre-release evaluation of GPT-6.2 Astra on four AI R&D tasks. It scores 29.01% on the RSI-Index, nears human references in compression and LM training, completes day-long tasks unaided, and underperforms on open-ended MoE post-training.
Why it mattersPre-release testing shows GPT-6.2 Astra nearly matches human references on pre-training tasks (within 1-3%) but falls below baseline on open-ended post-training.
Anthropic's ClaudeDevs account announces linked billing in beta on the Claude Platform. Users of integrated tools such as Factory, Linear, and Ramp can pay for Claude API usage via their own organization's account, with no API key shared. Setup details were not in the post.
Why it mattersLinked billing in beta lets users of tools like Factory, Linear, and Ramp pay for Claude API usage through their own Claude Platform organization. Tool builders no longer need users to hand over an API key.

Vercel Blob now supports TTLs for automatic deletion
Vercel Blob now supports automatic TTL-based deletion (1-730 days) on uploads, copies, and presigned URLs.
Why it mattersRemoves the need to build and run your own cleanup job for temporary exports, session transcripts, or uploads stored in Vercel Blob.
Artificial Analysis evaluates OpenAI's GPT-6.2 Astra and Sol: #2 on its Intelligence Index behind Fable 5.2, 20% cheaper Astra pricing, dominant cost-per-task frontier, but no gain on the Coding Agent Index and 55-59% hallucination rates.
Why it mattersGPT-6.2 Astra is priced 20% lower ($8/$40) and ranks #2 on the Intelligence Index, but coding-agent scores are flat (Astra) or down 3 points (Sol). Engineers choosing coding models should weigh cost gains against no coding improvement.

GPT-6.2 Astra and GPT-6.2 Sol—Models built for the real world
OpenAI releases GPT-6.2 Astra for the hardest projects and GPT-6.2 Sol for everyday work.
Why it mattersTwo new OpenAI models are available in Codex and the API as gpt-6.2-astra and gpt-6.2-sol. They claim better results with fewer output tokens, which affects cost and usage limits for agent workloads.
OpenAI announces GPT-6.2 Astra and Sol as now available, pitched as efficient and built for real-world work. The accompanying demo shows spreadsheet and presentation tasks. The full release article was unreadable, so specs, pricing, and benchmarks are not covered here.
Why it mattersOpenAI released GPT-6.2 Astra and Sol, two new models now available to users. Engineers choosing models should check the official release notes for capabilities, pricing, and limits, which this post does not detail.

How Close Is Self-Improving AI? OpenAI & Thinking Machines Researchers Explain
a16z's Working Theory episode with Mark Chen of OpenAI and John Schulman of Thinking Machines on recursive self-improvement.
Why it mattersHear OpenAI and Thinking Machines researchers on how far AI automation of AI research has progressed, including OpenAI's AI research interns, and why research taste and alignment remain hard.
swyx reports a week using Devin bots, always-on cloud agents that own outcomes, manage subagents and report via interactive artifacts. He argues for outcome-owning agent managers over 'AI employees'. Product details come from a conference preview and include a promotional credits note.
Why it mattersDescribes Devin bots as persistent cloud agents that own outcomes and manage subagents, with a week of hands-on use across QA, porting and automation.

Co-Designing Sparse Attention for Efficient Long-Context Inference
NVIDIA details how to co-design sparse-attention indexers for long-context inference.
Why it mattersConcrete architectural choices (indexer sizing, fused Top-k, selection reuse) that directly cut the cost of serving long-context LLMs, useful for anyone building or tuning inference infrastructure.

Introduction to GPU Inference, a short course with Interactive diagrams
A free, signup-less course of 5-10 minute narrated lessons with interactive diagrams on GPU inference, progressing from basics to production serving concepts like continuous batching, goodput.
Why it mattersExplains how LLM inference is actually served, including continuous batching and disaggregated prefill and decode, which drive the cost and latency limits agent builders run into.
Engineering AI Discover
Netflix engineers explain the intelligence layer behind AI Discover.
Why it mattersShows how a large consumer product structured an LLM planning layer on top of existing recommenders while keeping TV latency, reliability and safety constraints. Useful reference for production agent harness design.

Put Gemini to work for you
Google adds Task mode to the Gemini app so it can schedule appointments, analyze bills and buy items through connected services, executing inside isolated virtual machines and reachable via app toggle, text or email.
Why it mattersGemini now has a Task mode that acts on email, calendar and payment methods, running in hardware-isolated virtual machines. It shows how a major lab scopes delegated agent actions.

Inference engineering for software engineers
The Pragmatic Engineer examines how engineering teams move from frontier lab APIs to open models.
Why it mattersCompanies report cutting per-token costs 50 to 90% by moving to open models. This explains the inference-provider path and the operational issues to expect: tail latency, cache hit rates and rate limits.
Cloudflare Joins the Jevolution and Labels Itself a Neolab
Latent Space reports that Cloudflare launched Clef, a decision model, and Clairic, an extract model, and calls itself a neolab.
Why it mattersCloudflare is moving from hosting models to shipping its own small 'system one' decision and extract models. If you run agents on its platform, expect first-party low-latency models rather than only third-party ones.

unsloth v0.1.919-beta — Computer Use + Translate + MiniMax-H3-Image + New Model Picker
Unsloth adds beta computer-use for local vision models, MCP integrations with Unity/Unreal/Blender/Roblox, GGUF/FP8/INT8 and LoRA support for MiniMax-H3-Image.
Why it mattersBrings local-model computer-use and game-engine MCP control to a widely used fine-tuning/inference tool, useful for anyone building local agentic workflows outside the cloud.

What is the best harness for each model?
RUNTA and Baseten benchmarked 22 coding-agent harnesses across 5 models and 62 tasks (27,280 runs), finding the best-value harness differs per model and cost ranks shift sharply when you switch models.
Why it mattersIf you switch models, your harness choice may need to change too: Codex CLI topped cost-efficiency with GLM-5.3 but ranked 16th with Kimi-K3, a concrete, decision-relevant finding for anyone running coding agents.

open-webui v0.13.4
Open WebUI's Tether companion exposes local models, MCP/tool servers, and terminals from your own machine into any connected Open WebUI instance.
Why it mattersLets you run local models and coding agents through Open WebUI from any device while keeping credentials and file access confined to your own machine, with sandboxing instead of Docker/VM setup.

We don't need more apps. We need better open source
Jordan Krueger argues that AI-assisted coding should target existing open source: faster performance, one-command installs, easier operation and lighter resource use.
Why it mattersArgues that AI-assisted coding is better spent on speeding up, simplifying and operationalizing existing open source than on new apps. It names four concrete jobs, with uv as the example.

Ian Buck Kicks Off NVIDIA’s Week at SC26 as Science Enters the Age of Agents
NVIDIA's Ian Buck announced Nemotron for Science (an open model runnable on a DGX Spark with a new chemistry dataset), the Rubin HPC GPU, an expanded Vera CPU platform.
Why it mattersA new open science-focused model you can run locally plus new HPC hardware, relevant if you're building AI-assisted research or simulation tooling.

Qwen3.8 Flash Next NVFP4 QAD released
NVIDIA published an experimental QAD checkpoint of Qwen3.8-Flash-Next using NVFP4 for experts and MXFP8 for attention and GDN projections.
Why it mattersIt documents which components of a large MoE model can be held in NVFP4 versus MXFP8 and the resulting ~130 GB footprint, useful if you plan low-precision serving on Blackwell-class hardware.

opencode v1.19.6
OpenCode v1.19.6 adds an optional upgrade to OpenCode 2.0 (background download, choice of restart, upgrade on quit, or later) and fixes oversized requests.
Why it mattersOpenCode 1 now prompts to upgrade to 2.0, keeping sessions and settings, but plugins built for 1 need updates. Long image-heavy sessions also compact and retry instead of failing on oversized requests.
An index of the vibe-coding frontier. Corrections welcome.