Hermes Kanban: Mission Control for your Agents
AI security is often mistakenly focused solely on the application layer, such as prompt filtering and output guardrails. However, true security requires infrastructure-level controls to manage data boundaries, retention policies, and storage. AWS provides this through IAM policies, VPC isolation, and CloudTrail logging, allowing AI features to inherit robust security protocols by default. This approach is demonstrated by IDEMIA, which manages government identity data on AWS while significantly reducing transformation time. The Hermes Kanban project demonstrates a collaborative four-agent software development team coordinated through Telegram and a shared Kanban board. The team consists of specialized agents for project management, backend development, frontend development, and testing, who communicate by passing task summaries. To overcome the backend agent's tendency to exhaust its context window while managing infrastructure, the developers integrated InsForge as a backend context engineering layer. This approach successfully enabled the autonomous creation of a functional Google Docs clone with integrated AI features and real TypeScript edge functions. This section details twelve CLI integrations that expand the functionality of Claude Code for AI agents. These tools allow agents to perform complex tasks such as managing GitHub repositories, executing payments via Stripe, and fine-tuning models on HuggingFace. Key infrastructure tools like E2B provide secure sandboxes for code execution, while Unsloth optimizes local model training. Together, these integrations enable agents to operate more like human engineers within a terminal environment.
閱讀原文 ↗目錄
AI security has nothing to do with AI
AI security is often mistakenly focused solely on the application layer, such as prompt filtering and output guardrails. However, true security requires infrastructure-level controls to manage data boundaries, retention policies, and storage. AWS provides this through IAM policies, VPC isolation, and CloudTrail logging, allowing AI features to inherit robust security protocols by default. This approach is demonstrated by IDEMIA, which manages government identity data on AWS while significantly reducing transformation time.
- Application-layer security like prompt filtering does not control where data travels or how it is stored.
- Regulated data can leave organizational boundaries when hitting external model provider endpoints.
- AWS infrastructure security includes IAM policies for service scoping and VPC isolation for tenant traffic.
- Encryption and CloudTrail logging cover all data flows, including model calls, by default on AWS.
- IDEMIA achieved a 4x reduction in transformation time by leveraging AWS infrastructure security for government data.
- AI features inherit existing infrastructure security protocols from the first day of implementation.
Hermes Kanban: Mission control for your agents
The Hermes Kanban project demonstrates a collaborative four-agent software development team coordinated through Telegram and a shared Kanban board. The team consists of specialized agents for project management, backend development, frontend development, and testing, who communicate by passing task summaries. To overcome the backend agent's tendency to exhaust its context window while managing infrastructure, the developers integrated InsForge as a backend context engineering layer. This approach successfully enabled the autonomous creation of a functional Google Docs clone with integrated AI features and real TypeScript edge functions.
- The multi-agent system uses a shared Kanban board to maintain state and facilitate handoffs between specialized agents.
- Backend engineering agents often fail when building infrastructure from scratch due to context window exhaustion and inconsistent state management.
- InsForge serves as an open-source, agent-native backend layer that provides agents with structured 'skills' for infrastructure tasks.
- A 'skill' is defined as a step-by-step guide that prevents agents from improvising infrastructure and ensures reliable deployment.
- The system successfully built a Google Docs clone, managing database tables, user authentication, and document handling autonomously.
- Communication between agents is handled via task summaries that define API shapes and verification requirements for the next agent in the workflow.
The top Claude Code CLI integrations
This section details twelve CLI integrations that expand the functionality of Claude Code for AI agents. These tools allow agents to perform complex tasks such as managing GitHub repositories, executing payments via Stripe, and fine-tuning models on HuggingFace. Key infrastructure tools like E2B provide secure sandboxes for code execution, while Unsloth optimizes local model training. Together, these integrations enable agents to operate more like human engineers within a terminal environment.
- GitHub integration enables agents to interact with issues, PRs, and Actions rather than just editing text.
- Bright Data provides automated unblocking and proxy support for web scraping via terminal prompts.
- InsForge consolidates backend services like databases, auth, and hosting into a single CLI tool.
- E2B creates isolated VM sandboxes to safely execute and capture output from agent-written code.
- Unsloth speeds up LoRA and QLoRA fine-tuning by 2x while significantly reducing VRAM usage.
- Playwright allows agents to interact with live web pages and perform UI testing across multiple browsers.