Great! Hermes Agent v0.13.0 'Tough Version' is here! I am building an integrated architecture solution for Hermes+DeepSeek+Vibe Trading system to achieve an AI automated trading system! Please see the system architecture diagram of Hermes+DeepSeek+Vibe Trading that I am currently designing, as shown in Figure 1. On May 7, 2026, Hermes Agent officially released version 0.13.0, codenamed Tenacity "Tough Edition", allowing AI to go from "three minute heat" to "reliable and capable"! Refer to Figure 2 and the video. The 'Tough Version' uses a hardcore update to completely solve the old problems of AI such as' incomplete, off track, and chaotic collaboration '. This is not a minor fix, but a "low-level refactoring" that gathers 295 contributors, 864 code submissions, and adds over 120000 lines of code within a week, with only one goal: to truly teach AI how to get things done and done well! Let's take a look at the new version of hardcore technology and highlights together. Hermes Agent's latest v0.13.0 tough version focuses on underlying technology refactoring, with a core of hardcore capability iteration and no flashy features. 1. Breakthrough in the Core of Multi Agent Collaborative Kanban Built in heartbeat detection, zombie instance recycling, and illusion gate control mechanisms: automatic transfer of stuck tasks, real-time interception and error correction of AI fabricated content; The entire task is persistent, with no loss of progress during gateway restarts or program crashes, completely solving the pain points of long task abandonment, interruption, and restart. Launch a dedicated AI task dashboard system that supports parallel division of labor among multiple agents and automatic claim and circulation of tasks. 2. Persistent target locking architecture New/goal locking mechanism has been added, relying on Ralph loop logic to permanently anchor core tasks. In multi round conversations and complex branching interactions, AI will not deviate from the initial requirements, curing the problem of traditional agents forgetting their original intention while chatting. 3. Native video understanding multimodal capability Built in video_analyze dedicated toolchain, directly connected to multimodal large models such as Gemini, native support for video parsing, automatic disassembly of video content, extraction of key information, recognition of image logic, no longer relying on subtitle to text indirect analysis. 4. Bottom layer reinforcement for global security Strengthen cross platform permission isolation, block Discord and social channel permission leakage vulnerabilities, comprehensively upgrade OAuth and identity authentication, and safeguard data security from the bottom. Fix 8 high-risk security vulnerabilities at once, automatically desensitize sensitive information by default, mask keys and privacy credentials; 5. Session self-healing and architecture optimization Refactoring the session breakpoint continuation mechanism to achieve automatic recovery of task and conversation states; Add Google Chat access, compatible with 20 mainstream platforms; Large scale optimization of the underlying code, co built and submitted by nearly 300 developers in the community, greatly improving runtime stability and plugin extensibility. Overall, this version is not just about adding features, but rather completing the hard core transition of AI agents from "usable" to "stable and easy to use" from four dimensions: multi-agent collaboration, target control, multimodal analysis, and security architecture.