Full DAG test episode

Cloudflare's MCP server is the reference example — two tools, search and execute, cover roughly 2,500 endpoints in about 1K tokens.

• Three paths to connect agents to systems: direct API, CLI, and MCP — mature integrations will ship all three, but MCP is the compounding layer for cloud-hosted agents.
• MCP SDK downloads hit 300M/month, up from 100M at the start of the year.
• Group tools around intent, not endpoints — a single create_issue_from_thread beats four primitives.
• For huge surface areas (Cloudflare, AWS, Kubernetes), expose just search+execute and let the agent write code.
• Tool search cuts tool-definition tokens by 85%+; programmatic tool calling cuts multi-step token usage by ~37%.
• Skills and MCP are complementary: MCP gives raw tool access, skills teach the procedural 'how.'

Run `/ultrareview` before merging critical changes — auth, data migrations, etc. Pro and Max users get 3 free reviews through 5/5.

• New /ultrareview command runs a fleet of bug-hunting agents in the cloud, findings land in the CLI or Desktop.
• After the three free reviews, each run is billed to extra usage at $5 to $20 depending on change size.
• Early reports in the replies describe failed runs that still consume the allowance.
• Positions Claude Code against specialist PR review tools like Greptile, CodeRabbit, and Devon.

I'm back to the drawing board, because the budget I thought I would need is blown away already.

• Uber's CTO said the 2026 AI budget was blown in the first four months, largely due to Claude Code uptake.
• About 1,800 code changes a week at Uber are now written entirely by its internal AI agent with no human input.
• ~95% of Uber engineers use AI tools monthly; ~70% of committed code is AI-generated.
• Internal AI agent share of code changes grew from under 1% to about 8% in a few months.
• Anthropic's enterprise shift to per-seat + mandatory token commits removes the predictability of old fixed-tier billing.
• New internal phenomenon: 'tokenmaxxing' — ranking engineers by token spend (reported at Meta).

Running too many agents in parallel that don't communicate with each other is almost useless compared to fewer agents burning tokens efficiently in a critique loop.

• December 2025 was the phase transition: Shopify's internal AI tool usage went from steady growth to runaway, now approaching 100% daily active.
• CLI-based agents (Claude Code, Codex, Shopify's internal 'Roast') are outgrowing IDE tools like Cursor and Copilot.
• Token distribution is skewing — top-percentile engineers consume disproportionately more, which Parakhin calls 'not ideal.'
• Parakhin defends Jensen Huang's 'engineer should spend $100K in tokens' framing as 'directionally correct' but warns quantity-without-critique is an anti-pattern.
• Shopify uses Graphite stacks and has not found an external PR review tool that runs pro-tier models with the right critique loop — they built their own.
• Merge conflicts are becoming a global mutex problem as code volume outpaces human review throughput.

If an error case is rare, pessimize the error case and optimize the success case. Don't allocate unnecessarily on the happy path if it's going to succeed 99+% of the time.

• Agents (and humans) default to tracking state for undo, which allocates on the hot path even when the error never happens.
• Better pattern: on error, re-run from the start in a slow path that does the undo work — Zig's parser uses this for error info.
• Conditionals are nearly free when predictable; memory access is not, even cached, because cache locality dominates.
• Hashimoto doesn't expect models to reason to this pattern generically in the short term — tool design and result criteria have to push for it.

Shared agents that can handle complex tasks and long-running workflows across tools and teams.

• Workspace agents: build once, share across a team; can pull context from docs, email, chat, code.
• Can take approved actions in Linear, Slack, and other connected systems.
• Research preview for ChatGPT Business, Enterprise, Edu, and Teachers plans — not on Plus.
• Positioned as the 'shared agent' layer, with OpenAI explicitly pitching cross-team workflow automation, not individual assistants.

Traditional computer vision tasks — segmentation, depth estimation, normal prediction — can now be performed at or near SOTA with a single generalist model.

• Vision Banana is built on Nano Banana Pro; the claim is that image generation is a universal interface for vision tasks.
• Segmentation, depth estimation, and surface-normal prediction can be framed as image-to-image generation and hit near-SOTA with one model.
• Project page at vision-banana.github.io; paper on arXiv (2604.20329).
• Sits alongside OpenAI's GPT Image 2 (which topped LM Arena by a record 242 points last week) as evidence the image-gen stack is absorbing classical CV.

The new Qwen3.6-27B just gave me definitely the best pelican riding a bicycle I've had from a 16.8GB model file.

• Qwen3.6-27B dense: 16.8GB quantized, runs on a MacBook.
• Alibaba Qwen team claims the 27B dense surpasses Qwen3.5-397B-A17B MoE on major coding benchmarks.
• Unsloth GGUF builds landed on Hugging Face the same day.
• Willison's 'pelican riding a bicycle' SVG test has become an informal shorthand for small-model capability.

For every 10-percentage-point increase in observed exposure, perceived job threat increased by 1.3 percentage points.

• 1/5 of the 81K Claude-user respondents voiced concern about economic displacement from AI.
• Early-career respondents were much more likely to express job-displacement concern than senior workers — only 60% of early-career said they personally benefit, vs 80% of senior.
• Perceived job threat and speedup are U-shaped: people slowed down by AI, and people sped up most, both worry more than the middle.
• Highest- and lowest-paid occupations reported the largest productivity gains, with 48% citing scope expansion and 40% speed.
• Respondents overwhelmingly named themselves as the beneficiary of gains; 10% named employers, smaller shares named AI companies.

OpenAI just released a free healthcare version of ChatGPT-5.4 for clinicians that beat specialty-matched physicians with unlimited time and web access on a benchmark of real and hard clinical tasks. Caveat: the benchmark was designed by OpenAI, though it is fully open.

• OpenAI released a free clinician-facing build of ChatGPT-5.4.
• Claims to beat specialty-matched physicians who had unlimited time and web access on a benchmark of real clinical tasks.
• Benchmark was designed by OpenAI (Mollick flags the conflict) but is fully open, so third parties can rerun it.
• Mollick's framing is deliberately qualified, which is what makes the signal readable.

AI tools must not be used to generate, extract, or summarize material that is then attributed to a named source, whether as a direct quote, a paraphrase, or a characterization of someone's views.

• Reporters must not use AI to generate or summarize material that is then attributed to a named source.
• No AI-generated images, audio, or video published as authentic documentation of real events.
• Every author using AI in reporting must disclose the use to editors; author remains responsible.
• Policy frames accountability as non-transferable — can't hand off blame to an editor or a tool.

It's a bidirectional token-classification model for personally identifiable information detection and masking in text.

• OpenAI released 'Privacy Filter,' their first open model of 2026, under Apache-2.0.
• Adapts GPT-OSS into a bidirectional token-classification model for PII detection and masking.
• 1.5B parameters total, 50M active — small enough to run in a browser.
• Retains a 128K context window at that size, which Elie Bakouch called 'quite impressive.'

The key design principle behind the Claude Agent SDK is to give your agents a computer, allowing them to work like humans do.

• Claude Code SDK is being renamed Claude Agent SDK — same harness, broader positioning.
• Core pattern: gather context → take action → verify work → repeat.
• Subagents get first-class support for parallelization and context isolation.
• Code generation is the preferred action when a task needs precise, composable, reusable output.
• Three feedback patterns for verification: rules (linting), visual (screenshots), LLM-as-judge.

Sony built 'Ace,' an autonomous ping-pong robot that uses RL and Sony's vision sensors to achieve expert-level play. A huge leap forward for adaptive robotics.

• Published in Nature today by Sony.
• Ace uses reinforcement learning paired with Sony's custom vision sensors.
• Claims expert-level ping-pong play against humans.
• Sits on the trajectory from DeepMind's ping-pong work two years ago to full-match autonomous play.