AI & Automation Briefing - October 15, 2026
Figma Ships Agent Controls and Code Layers to Reshape Design-to-Dev Handoff
Figma announced a set of agent-focused features at its first Config India event that will directly affect how automation specialists connect design systems to development pipelines. The centerpiece is Figma Specs, a Markdown-based editor where teams define their design language, reference canvas elements, embed FigJam diagrams, and link design and code libraries. Both human teams and AI agents can query Specs as a centralized source of truth, giving builders a structured way to enforce brand rules across agentic workflows. Figma also moved its code layers feature from private to public beta, letting users generate and iterate on code directly from canvas designs. Agents can convert a design into code or modify existing code on demand, tightening the loop between prototyping and implementation. The platform's MCP now supports over 20 tools, including Open Code, Dia, Grok Bot, and Indian AI startup Sarvam. Separately, Figma integrated its acquired headless CMS Payload into Figma Make for generating dynamic websites, noting that Payload downloads have grown sixfold over the past year.
Most Agentic RL Self-Distillation Methods Fail to Beat a Simple Baseline
A large-scale benchmark study tested 10 self-distillation methods for agentic reinforcement learning against GRPO, a standard baseline, across four environments and three model sizes. The research logged 132 training runs totaling over 12,000 GPU-hours. The findings are sobering: no self-distillation method consistently outperformed GRPO across settings, strong final task scores often concealed unstable training behavior, and higher computational cost did not reliably produce better results. For teams evaluating or building AI agents, this study is a useful check against methods that look good on benchmarks but may not hold up under controlled, fair comparison.
Apate Deploys AI Agents as Decoy Scam Victims to Extract Fraud Intelligence
Australian startup Apate.ai, founded in 2024 out of Macquarie University research, runs thousands of AI agents that pose as potential scam victims. The agents answer fraudulent calls, respond to phishing messages, follow malicious links, and infiltrate social media groups, maintaining conversations long enough to extract details scammers would normally only reveal after a real victim is already compromised. That data includes bank accounts, crypto wallets, phone numbers, malicious domains, and impersonated brands. Apate sells this intelligence to banks, telecoms, government agencies, and utilities to strengthen their anti-fraud programs. To date, the company reports its agents have conducted 2.5 million conversations and consumed 17 years of scammer time. The scale of the underlying problem is significant: the FBI logged over 22,000 AI-related crime complaints in 2025, resulting in roughly $893 million in losses, with more than $632 million tied to investment scams run using AI-generated conversations and deepfakes. The concept traces back to a 44-minute scam call that Apate founder and CEO Dali Kaafar deliberately prolonged at a family picnic, which prompted him to ask whether the process could be automated at scale.