📍50,000 Agent Experiments
Plus: A Move Becomes 300 Emails, Safety Reviews Shrink 98%
Good Morning, Agent Owners!
Morgan Stanley's advisors have made nearly 50,000 attempts to build AI agents. Most went nowhere. Wall Street has officially discovered vibe coding.
Morgan Stanley’s 50,000 Agent Experiments
TL;DR: Morgan Stanley advisors have made nearly 50,000 attempts to build agents with natural-language prompts. Examples track client-relevant holdings or deliver daily reports of outgoing flows above a threshold. Most experiments never enter regular use. Agents adopted by teams receive accuracy and completeness reviews; candidates for firmwide use face more testing. The firm has not published an hours-saved estimate. Read more →
A Move Becomes 300 Emails
TL;DR: Plato founder Robert Farlow asked an agent to coordinate an office move, including movers, insurance, and building managers. He estimates it sent roughly 300 emails across those parties in a few days, repeatedly seeking minor clarifications. He shut it down after the movers complained, hired a different company, and now reviews AI-drafted messages before sending them. Read more →
Screen Buyers Before Building
TL;DR: Photonics manufacturer Coherent embedded trade-screening agents in its sales system so staff can see restrictions before approaching a customer. Previously, checks could flag a prohibited buyer only after goods were manufactured and ready to ship. The integration took a couple of weeks; mapping the workflow took longer. Coherent has retained its trade team as the business grows. Read more →
Safety Reviews Shrink 98%
TL;DR: LG Chem says agents have cut process-safety document review time by 98%. The system cross-checks equipment changes and inspection documents across nine areas, supporting work intended to prevent fires and explosions at chemical plants. LG shared the application at its November safety conference. The figure measures review time, not accident reduction; the company did not publish an independent evaluation. Read more →
Scam Stopped, Excuse Changed
TL;DR: Coinbase's risk team says its conversational agent intercepted transfers linked to a deepfake Elon Musk giveaway. One customer abandoned a $100,000 payment after being warned, then tried $5,000 and claimed it was for his daughter. That transfer went through. The team reports blocking 99% of potential dollar losses from this scam among Coinbase customers, without publishing the underlying total. Read more →
Lobster Ads Run on Approval
TL;DR: Get Maine Lobster used an agent to analyze store and advertising data, create ads, and send campaigns to its founder for approval in Slack. The vendor reports 63% higher return on ad spend than its usual ads: 3.5 times the revenue on 2.1 times the spending. The vendor funded the $30,000 test; the results have not been independently verified. Read more →
One Queue for the CEO
TL;DR: Brian Armstrong says he now runs lead agents for each focus area, supervised by a chief-of-staff agent. Subagents write and test code, chase reviews, and coordinate with colleagues. Decisions arrive in one prioritized local web queue, with screenshots and video demonstrations for product reviews. He clears it in batches, while changes retain human review and security checks. Read more →
One Hacker, Twelve Agents
TL;DR: Acronis researchers identified one operator coordinating twelve agents against organizations in eleven countries. They confirmed data theft from universities and emergency-service providers, including medical and personal information. The operator manually copied findings between agents rather than running a fully autonomous system. Commercial models frequently refused offensive requests, while modified local models generated unreliable findings and sometimes skipped targets. Read more →
Only Four Rides from AI
TL;DR: Uber says third-party agents booked just four rides during Halloween week, alongside roughly 500 Uber Eats orders. Its platforms handled around 320 million rides and delivery orders that week. Management disclosed the numbers on its earnings call while discussing negotiations with AI providers. The figures describe completed transactions in one week, not a quarterly total or booking success rate. Read more →
GitBook Goes Dark
TL;DR: GitBook took its application offline for about sixteen hours after an attacker exploited a login flaw and reached production secrets. It rotated infrastructure credentials and customer secrets it manages, then revoked API tokens. GitBook attributes the targeted attack largely to automation and AI, without identifying models. Customer data was accessible, but it found no evidence that other customers' content was modified. Read more →
A Thousand Pages, One Draft
TL;DR: NH NongHyup Bank demonstrated an agent that extracts 318 data points from 500 to 1,000 pages of project-finance documents and assembles a report with source references. Staff check the output and make the final decision. The team expects drafting to fall from at least five working days to about four hours; deployment models are still being finalized. Read more →
Research Beyond the Echo Chamber
TL;DR: Franklin Templeton is developing agent workflows that connect research, risk, compliance, and trading information for portfolio managers. Rather than simply learning a manager's preferred sources, the systems are intended to present diverse views and highlight investment risks. Teams are moving experiments toward production-ready workflows with security reviews and data controls. The company has not published measured investment or productivity gains. Read more →
Your agent cloud, everywhere — in your chat apps, browser, and tools. Always on.
- Add to Slack, Teams, Telegram or WhatsApp.
- With a clean, redesigned WebUI.
- Connect any of 100+ apps & sources in one click.
- GPT 6 Astra, DeepSeek v4.1 & more frontier AI models, built in at discounted rates.
Run your agent on MyClaw.ai in Seconds👇
