Agentic AI

Figma has implemented AI agents to assist its security team with alert investigation, historical incident search, system checks, and code fix preparation. These agents leverage learnings from previous investigations to reduce repetitive manual labor. The result is a significant speedup, with complex alerts being resolved approximately 70% faster than before.

• AI agents automate repetitive security tasks like alert triage and system checks.
• Resolution time for complex security alerts dropped by roughly 70%.
• Agents improve by learning from historical investigation data over time.
• Security teams can offload code fix preparation to AI assistance.

ByteDance released DeerFlow 2.0, an open-source framework designed to orchestrate long-horizon tasks through sub-agents, memory, and sandboxes. The tool aims to handle complex workflows spanning minutes to hours by leveraging extensible skills and a message gateway. It reached the top of GitHub Trending shortly after its launch, signaling strong community interest in autonomous agent orchestration.

• DeerFlow orchestrates sub-agents and memory for tasks lasting minutes to hours.
• Extensible skills and sandboxes enable code, research, and content generation.
• Reaches #1 GitHub Trending, indicating high demand for open super-agents.
• Provides a message gateway for sub-agent communication and state management.

Zhou Yu explains why AI agents often stall during the demonstration phase and introduces simulation-driven testing to overcome compliance and reliability hurdles. The approach utilizes synthetic user personas and trajectory entropy to rigorously evaluate multi-turn interactions. This method enables engineering teams to identify edge cases before deployment and scale self-learning workflows effectively in production environments.

• Simulation-driven testing resolves reliability and compliance bottlenecks common in AI agent demos.
• Synthetic user personas provide realistic scenarios for evaluating multi-turn agent behaviors.
• Trajectory entropy serves as a metric to detect and catch edge cases before deployment.
• Automated CI/CD pipelines are essential for scaling self-learning workflows in production.

Google has open-sourced Mantis, an AI-agent framework that automates the entire vulnerability lifecycle including identification, validation, reproduction, and remediation. The tool addresses the critical issue of high false positive rates and hallucinated findings common in standard AI-powered code scanners. By adding agentic reasoning, Mantis aims to deliver more accurate and actionable security insights for developers.

• Open-sourced by Google to tackle AI code scanning noise and hallucinations
• Automates full lifecycle: identify, validate, reproduce, and fix vulnerabilities
• Designed to significantly reduce false positives in static analysis tools
• Introduces agentic workflows for more reliable security validation

Context Mode addresses AI coding agent context exhaustion by sandboxing tool outputs, achieving a 98% reduction in token usage. The tool persists session memory and enforces routing across 17 platforms using MCP and hooks. This prevents agents from wasting context on filler text or forgetting file states during conversation compaction.

• Sandboxing tool outputs reduces context usage by 98%, preserving space for critical data.
• Persists session memory to prevent agents from losing track of editing tasks.
• Enforces routing across 17 platforms via MCP + hooks for structured agent interaction.
• Eliminates filler tokens and verbose explanations to maximize effective context window.

Camofox-browser is a server-based, stealth headless browser built on a Firefox fork with C++ level fingerprint spoofing. It serves as a drop-in replacement for Puppeteer and Playwright, specifically designed to bypass Cloudflare and anti-scraping measures. The tool enables AI agents to interact with the real web without triggering bot detection systems.

• Provides C++ level fingerprint spoofing to evade modern anti-bot systems like Cloudflare.
• Acts as a drop-in replacement for Puppeteer and Playwright scripts with minimal changes.
• Optimized for AI agents requiring reliable, stealthy web browsing capabilities.
• Runs as a local server on port 9377 for easy integration into agent workflows.