Today's GitHub trending tells a story about agents growing up: the emphasis has shifted from what a model can reason about to what an agent can actually do — interact with your files, remember context across sessions, and research the internet on your behalf. Alongside that operational maturity, there's a transparency subplot: a repo cataloguing leaked system prompts from major AI providers is climbing fast, reflecting a broader community interest in understanding how these systems are actually instructed.
OfficeCLI: AI Agents That Work With Your Office Documents
iOfficeAI/OfficeCLI picked up over 1,700 stars today, offering a C#-based suite that lets AI agents read, create, and manipulate Office documents without requiring Microsoft Office to be installed. The value here is practical: enterprise workflows are deeply tangled with Word, Excel, and PowerPoint, and an agent that can fluently work in those formats — generating reports, processing spreadsheets, or updating presentations — closes a meaningful gap between what agents can do in a sandbox and what they can do in an actual workplace. The "no installation required" aspect makes deployment significantly simpler across corporate environments where IT controls what software is present.
github.com/iOfficeAI/OfficeCLI
System Prompt Leaks: A Transparency Catalogue Gaining Traction
asgeirtj/system_prompts_leaks crossed 1,200 stars gained today with a straightforward premise: collecting and publishing extracted system prompts from major AI assistants and products. It's a project that sits at an interesting intersection of security research and AI transparency — system prompts define how a model behaves, what topics it avoids, and what identity it's told to adopt. Many of these have been extracted through jailbreaks or accidental disclosures over time. Whether you view it as a security concern or a public-interest transparency effort, the traction it's getting signals that the developer community wants to better understand the instructions shaping the AI tools they and their users interact with every day.
github.com/asgeirtj/system_prompts_leaks
TencentDB-Agent-Memory: Local Long-Term Memory for AI Agents
TencentCloud/TencentDB-Agent-Memory added over 300 stars today with a focused solution to one of the more persistent agent problems: memory. The TypeScript library implements a 4-tier memory pipeline — working memory, episodic memory, semantic memory, and a compression layer — entirely locally, with zero dependency on external APIs. Long-term memory that doesn't require sending your conversation history to a third-party service is a meaningful privacy and latency win for teams building production agents. The tiered approach also means agents can be selective about what they store and recall, rather than naively accumulating everything.
github.com/TencentCloud/TencentDB-Agent-Memory
Superpowers: A New Methodology for Agentic Software Development
obra/superpowers gained over 1,100 stars today with a Shell-based skills framework that positions itself as a new software development methodology built around agentic capabilities. The project introduces structured "superpowers" — composable skill units that AI coding agents can invoke — with the goal of making agent-assisted development feel less like prompting and more like programming. It's similar in spirit to addyosmani/agent-skills, but comes from a different angle: where Osmani's approach is a curated library, superpowers reads more like a manifesto with tooling, arguing that the right abstraction layer changes how developers think about what they're building and who (or what) they're building it with.