Martin Krasser

The detected constraints are written to concise, topic-scoped files that give agents initial guidance without overwhelming context. With these in place, agents produce code that matches existing patterns on the first pass instead of after course corrections. A more consistent codebase feeds back into more consistent agent output.

Concrete example from a recent pass on github.com/gradion-ai/fre…. The agent identified 8 constraint categories (type system conventions, config persistence patterns, async/sync boundaries, module dependency direction, widget factory signatures) and 7 consolidation opportunities. All implemented in a single session.

A pattern for making older repos more agent-ready. Use a coding agent to detect constraints, invariants, patterns, and symmetries at the architecture/design level. Detect exceptions and explore consolidation options. Let the agent fix exceptions and consolidate the codebase.

Blog post: gradion-ai.github.io/agents-nanny/2… Claude Code plugin: gradion-ai.github.io/ipybox/ccplugi…

🔧 Code actions invoke tools via executable code 💾 Working code actions are saved as reusable tools 🧩 Typed interfaces are separated from implementations 🚀 Agents can accumulate new capabilities at runtime 🧬 Tool libraries evolve based on agent experience ♻️ Reusable tools avoid redundant code action generation

New blog post and Claude Code plugin: Code Actions as Tools Programmatic tool calling becomes more powerful when code actions are saved as reusable tools, enabling agents to evolve their tool libraries at runtime.

@wergieluk It uses Anthropic's srt to sandbox the IPython kernel that executes code, and optionally individual MCP servers.

@wergieluk When running ipybox as an MCP server, I can also be started inside a Docker container.

A Claude Code plugin for programmatic tool calls. Instead of JSON, call MCP tools via generated Python APIs: 🛠️ Compose multiple tools with loops, conditionals, ... 📌 Keep intermediate results in variables, out of context 🧩 Full type hints for better tool chaining All code runs in a local sandbox (using srt).

Docs: - gradion-ai.github.io/ipybox/ccplugi… (👤) - gradion-ai.github.io/ipybox/ccplugi… (🤖)

Key features: 💾 Save successful code actions as reusable tools 📚 Build executable skill libraries over time ♻️ Stateful execution with variable persistence 🔌 Works with any MCP server (stdio, HTTP, SSE)

Replace arxiv -> quickarxiv on any paper URL to get a blog post version with figures and insights. Example: quickarxiv.org/abs/2511.10395

Adding group intelligence to existing AI systems is a trending topic. Here's a scalable approach using a group reasoner layer: gradion-ai.github.io/agents-nanny/2…

Read more on the trending topic of multi-party conversational AI at gradion-ai.github.io/agents-nanny/2…

We built this architecture to make existing AI agents work in group chats without modification. Check out: - Group Genie: gradion-ai.github.io/group-genie/ - Group Sense: gradion-ai.github.io/group-sense/ Open source. Framework agnostic. Ready to enable multi-party AI conversations.

Multi-Party Conversational AI 🧵 Most AI agents excel at 1-on-1 conversations but struggle in group chats. The problem? They're trained for direct queries, not multi-party exchanges where: - Context emerges across multiple participants - Information fragments across messages - People reference earlier points without repeating the full context Group conversations need a different architecture.

Project repository: github.com/yctimlin/mcp_e… Architecture: separate canvas server and MCP server components. Both production-ready.

TIL about mcp_excalidraw, an MCP server that connects AI agents to a live Excalidraw canvas for conversational creation of diagrams. Key features: - Real-time WebSocket sync between MCP server and canvas - Mermaid diagram conversion to Excalidraw elements - Available as local or Docker deployment Used it to generate diagrams from hand-drawn sketches. Results are impressive. Supports concurrent direct diagram manipulation.