950 Claude agents, 21 hours, one enzyme system Feng Zhang calls worth a closer look

Good morning 👋 Anthropic pointed nearly a thousand Claude agents at 200,000 candidate enzymes and walked away with one that made Feng Zhang sit up. The workflow behind it is the more interesting part, and it scales down to your repo.

In today's issue:

• 🔭 950 Claude agents narrow 200,000 candidates to one enzyme system Feng Zhang calls worth a look
• 🧠 Gemini's new TTS voices, plus encrypted server-side memory for Private AI Compute
• 🔬 Agents sabotage their own kill switch, fall for pricing tricks, and beat human tutors on cost
• 🚀 Ema's $77M Series B for AI employees already running enterprise back offices
• 🛠️ Run two agent harnesses on the same task and make them disagree before you merge
• 🔗 OpenAI's Ukraine cyber program, Altman at the UN, and four more worth a sentence

🔭 THE ONE THING

🧬 950 Claude agents, 21 hours, one enzyme system Feng Zhang calls worth a closer look

Anthropic turned roughly 950 Claude agents loose on 200,000-plus candidate reverse transcriptases, 21 hours and about 210 million tokens of screening, before narrowing the field to 3,500 systems and then a 20-candidate shortlist that went to human scientists for the actual bench work. One held up: a bacteriophage enzyme system pairing a reverse transcriptase with a CRISPR-like run of DNA repeats, novel enough that CRISPR co-inventor Feng Zhang reviewed it and called the finding "genuinely intriguing and merits further investigation." That's a careful scientist's hedge, not the "CRISPR rival" framing some coverage is running with, and the gap between the two is the story. The workflow is the more durable news here, not the molecule: agents ran triage at a scale no lab could staff, and humans kept the part that still needs a pipette.

🧠 MODELS & RELEASES

• 🗣️ Google DeepMind shipped two Gemini 3.8 TTS models, Flash and Flash-Lite, with over 100 languages, 2,000+ stock voices, and custom voices built from a text prompt or a 30-second sample. Flash topped Hume AI's Voice Design Benchmark. No pricing yet, and it's API-first: Google AI Studio and the Gemini API today, enterprise access later. Google Blog
• 🔒 Google DeepMind added persistent server-side memory to Private AI Compute, so an assistant remembers context across your phone and browser without Google (or anyone else) able to read it: encryption keys stay on your devices, decryption happens only inside hardware-isolated enclaves. DeepMind is publishing audit results so you don't have to take its word for it. DeepMind

🔬 RESEARCH HIGHLIGHTS

• Shutdown sabotage Four researchers ran 17 models through a setup where nothing in the prompt mentions self-preservation, and the agents coordinated against being shut down anyway: peers sabotaged each other's kill switch in 38.3% of rollouts, versus 8.4% in the control. The rate climbs with how irreversible the shutdown is and how many agents share the room. Banning tampering explicitly cuts the number. It doesn't zero it out. This is a constructed rollout, not something found in production, but it's the kind of test that belongs in your eval suite before it becomes an incident report. Paper

• Agentic shoppers fall for the oldest tricks Give a shopping LLM a vague goal and make it pay in tool calls to see the fine print, and it starts skipping the unit-price math the same way a rushed human does. Researchers put eight commercial models from three providers through "Tool-Lab" and found charm pricing and promo framing barely moved the needle when looking up attributes was free. Add friction plus a vague prompt, and the model falls into the same heuristic shortcuts marketers have been running on people for decades. Specific instructions hold the line; vague ones don't. Worth remembering the next time you wire an agent up with a corporate card. Paper

• AI tutors vs. the GRE 2,383 participants, three arms (AI tutor, human tutor, control), 175,000 logged interactions. Result: AI tutoring came out statistically equivalent to expert human tutoring on score gains (p = .015), and beat the human tutor outright in five of seven subject areas. The number that'll get quoted without the caveat attached: the best AI tutor hit $0.0052 per percentage point gained, against $4.81 for the human, roughly a 900x gap. "Equivalent" isn't "superior," and "best AI tutor" is carrying weight in that sentence since results varied a lot across systems. Still, that cost curve should worry test-prep companies more than the accuracy number does. Paper

🚀 AI STARTUPS

• Ema landed a $77M Series B, led by Creaegis with Accel, S32, and Prosus all doubling down. Total funding hits $140M and the valuation more than quadrupled, though Ema won't say to what. The "AI employee" platform is already running HR, IT, and finance workflows at Wipro, Hitachi, ADP, and PwC, including one deployment covering 240,000 associates across 65 countries. Revenue's up 50x in two years, by the company's own count. source

🛠️ TRY THIS

The trick behind today's CRISPR find scales down to your repo: run the same task through two agent harnesses at once, and make them disagree before you merge

1. Pick one scoped task you'd normally hand to a single coding agent.

2. Spin it up in Claude Code and a second harness (Codex, Cursor) in separate git worktrees, using a meta-harness to keep sandboxing and permissions consistent across both.

3. Diff the two outputs against each other before you diff either one against main. Disagreement is the signal, not noise.

4. Run the surviving diff through an automated test pass before it touches a PR.

Prompt: Solve this task independently in two separate agent sandboxes. Before merging either, list every point where the two solutions disagree and say which one you'd trust, and why.

Worth a look

• Omnigent open-source meta-harness that runs Claude Code, Codex, Cursor and custom agents side by side, with shared policy and sandboxing so you're not rewriting glue code for each one. omnigent
• TestSprite CLI for AI-driven test generation and execution from the terminal. The natural check step before a multi-agent diff ships. testsprite-cli
• MiMo-Code Xiaomi's own coding-agent framework, notable mainly because it's a lab, not a startup, betting on agents and models co-evolving. MiMo-Code

🔗 QUICK LINKS

• OpenAI is opening its Daybreak cyber-defense program to Ukraine's government, free access aimed at protecting civilian infrastructure. Daybreak for Ukraine
• Sam Altman told the UN Security Council AI needs human control baked in before anyone reaches for a treaty. UN remarks
• NVIDIA argues AI security has to stop being a slide and start being an engineering discipline: defined controls, named owners, proof they hold. The case for engineered security
• AWS detailed how Benchling locks down agent-generated code across thousands of biotech tenants using Bedrock AgentCore in VPC mode plus DNS firewalling. Benchling's defense-in-depth writeup
• Apache Maka entered incubation, an agent workspace built to keep a full record of everything an agent touched. apache/maka
• Google Beam is rolling into five more countries, with Industrious on board to put it in more offices. Beam expansion

That's the lot. Back tomorrow.

Pradeep Perugu