Claude Code goes hands-off by default on Aug 14

Good morning 👋 On August 14, Claude Code stops asking permission by default. Whether that's a productivity unlock or a liability depends on which number in today's lead you trust.

In today's issue:

• 🔭 Claude Code goes hands-off by default on Aug 14
• 🔬 RAG's blind spot on financial tables, plus video models that can't count what they're watching
• 🚀 OpenAI's quiet slide-deck buy, and an ex-Sora researcher's bet on plain language over prompts
• 🏛️ A polysilicon price floor aimed at Chinese supply
• 🛠️ Rate-limit your coding agent before auto mode does it for you
• 🔗 A tutoring eval, two AgentCore case studies, and three filings worth a glance

🔭 THE ONE THING

🤖 Claude Code goes hands-off by default on Aug 14

Starting August 14, auto mode becomes the default for every Claude Code user on Pro, Max, and Team plans, no opt-in required. Anthropic's own numbers: the safety classifier catches about 89% of dangerous commands against 13.6% for a human skimming the same diff, and teams running on auto mode shipped roughly 25% more pull requests. Trajectory Labs ran 720 independent prompt-injection attempts against Fable 5, Opus 5, and Sonnet 5 in auto mode and landed zero, and Anthropic isn't charging extra for the classifier's overhead either. I buy the productivity number; that's the kind of lift you'd expect from removing the "approve every tool call" tax. The 89% catch rate is the one to sit with: an 11% miss rate on dangerous commands is still roughly one in nine, and come Aug 14 it's running under everyone by default, not just the people who used to opt in.

🔬 RESEARCH HIGHLIGHTS

• RAG breaks on financial statements, and better embeddings won't save it. A new paper tests standard chunk-and-embed retrieval against financial and audit documents, where 87% of the content is table rows and figures routinely sit 13 lines from their own headers. On a 51-question benchmark built from a 780-page government financial report, dense retrieval scored 15.7%; a tuned version got to 35.3%. The authors' embedding-free system, which swaps similarity search for lexical search and structural navigation exposed as MCP tools, hit 58.8%. Plain BM25 matched it almost exactly: the real gap is embeddings versus no embeddings, not agentic versus simple. If you're running RAG over 10-Ks or audit reports, the fix is indexing structure, not fine-tuning your retriever. paper

• Can an LLM optimize its own harness? HarnessOpt-Bench put five frontier models in the optimizer seat, each iteratively rewriting a target agent's prompts, tools, and orchestration code against eval feedback, across 4 tasks and 111 scored runs. The model doing the optimizing swings results more than the harness it runs through, and a model's own native harness doesn't reliably beat a generic one. Worth knowing before you buy into the "auto-tuning agent" pitch: the capability is real and measurable, but there's a lot of room left before it's dependable. paper

• Video models can watch something happen and still not count it. Testing Gemini 3.6 Flash on 2,190 controlled clips with frame-level ground truth, researchers found it reliably counts persistent state changes up to 12 events, but has no reliable range at all for transient events like blinks. Push it into high-count, high-frequency territory and the final tally is right just 0.2% of the time, capturing only 18% of what actually happened. Feeding the model more frames boosted its raw accuracy on a bouncing-ball task from 19.6% to 29.3%, but the actual event sequence only matched ground truth 3.7% of the time, so the score went up while the model kept guessing. Don't hand a video model anything that needs an accurate count of fast, repeated events. paper

🚀 AI STARTUPS

• NextSlide OpenAI bought the presentation-slide startup back in early 2026 and sat on the news for months before founder Ahmed Beshry posted about it this week. The whole team is folded into ChatGPT's slide tools now. Terms undisclosed, which is how OpenAI handles these when the number's small. NextSlide joins OpenAI
• Energy Gabriel Petersson helped ship Sora 2, then walked away from OpenAI three months ago to build something plainer: an agent that runs your computer off natural-language instructions instead of a prompt-engineering ritual. His bet is that the models already cleared the bar. The interfaces didn't. Energy launch

🏛️ POLICY & REGULATION

• White House A new proclamation puts a floor under polysilicon prices: $21/kg for the base material, $100/kg for ingots and wafers, $0.22/watt for solar cells, $0.38/watt for modules. Undercut those floors and a tariff kicks in to close the gap, plus a flat 15% duty on polysilicon ingots and derivatives once it takes effect December 4. Japan, Korea, Taiwan, Switzerland, and the EU get carved out at the standard rate; the UK gets 10%. Cited authority is Section 232, the national-security route, which means this is dressed as chip and solar supply-chain protection even though the real target is Chinese-linked polysilicon capacity. Builder read: if silicon-grade material touches your hardware or solar supply chain, get your sourcing and pricing documentation sorted before December. Proclamation text

🛠️ TRY THIS

Put rate limits on your coding agent before auto mode does it for you

1. Claude Code's auto mode goes default for Pro, Max, and Team plans on August 14. Before then, list every irreversible action it could take unsupervised in your repos: deploys, prod credentials, external API calls, spend.

2. AWS just shipped a playbook worth stealing. AgentCore's new Dogwood policy language sets temporal and sequence rules on agent actions, and gateway rate limits cap cost no matter what a single session does. Read AWS's writeup. You don't need Bedrock to copy the pattern, just hard caps instead of soft prompts.

3. Wire usage into somewhere your team already looks. AWS's companion post routes Codex's OpenTelemetry metrics into CloudWatch, broken out by user, team, and cost center. Same idea works for any coding agent: pipe session and token data into whatever dashboard already pages someone.

4. Run one open-ended task through auto mode this week and confirm your cap actually fires before you trust it unattended.

Prompt: List every irreversible action this coding agent could take unsupervised in this repo (deploys, migrations, external calls, spend). For each, propose a rate limit or an approval gate.

🔗 QUICK LINKS

• AllenAI shipped TutorMoments, an eval that hands a model 462 real K-7 tutoring transcripts and checks whether it knows when to step back. Told to just "tutor well," most models tested over-helped instead. Hugging Face
• Cohere Health built its clinical-policy digitizer on Amazon Bedrock AgentCore, a multi-tenant, MicroVM-isolated agent stack that cut per-policy digitization time by 30%, from 2:15 to 1:35. AWS
• PDI Technologies turns plain-English requests into live, multi-tenant web apps with PDI Brew, a two-agent Bedrock deployer that skips Git, terminals, and DevOps entirely. AWS
• MARA Holdings, the bitcoin miner, filed its Q2 2026 shareholder letter with the SEC. Filing
• Innodata, which builds the data pipelines that feed frontier model training, filed its Q2 2026 earnings exhibit. Filing
• Cerence, the in-car voice-assistant vendor, filed Q3 FY26 earnings with the SEC. Filing

Set your own limits before auto mode sets them for you.

Pradeep Perugu