A lab that runs itself

The lab runs itself. Vlad sets the direction. Agents research, plan, build, test, review, and document the work. The system turns intent into tasks, routes them through tools and checks, and keeps enough evidence around the work that the next step can start from context instead of from scratch.

That is the point of iftodo. The website is the front door into the same operating model: agents, tool connectors, identity, retrieval, evaluation, local models, and the software around them. Paid work funds reusable lab infrastructure, and that infrastructure makes the next piece of work easier to run.

A model alone does not become a lab. It needs tasks, memory, source material, permissions, tools, tests, review, and a way to hand work from one actor to another. The interesting frontier is not a chatbot that sounds busy. It is an organization where pieces of work move through a system with enough context and traceability to be improved.

The projects here point at that shape. TrustChain asks how agents carry identity and signed history. MCP connectors turn real accounts and documents into capabilities. Retrieval work keeps answers tied to sources. Automated testing and model research make evaluation part of the loop. Local multimodal work keeps private media close to the machine.

In our view, the field is moving toward programmable organizations, agent networks, tool protocols, durable context, research automation, on-device AI, model adaptation, recommendations, preferences, and evaluation-driven systems. These are the parts of an agentic lab that has to know what it is doing, who is acting, what evidence exists, and when to stop.

iftodo starts from Delft and from built work. Some of it is public, like TrustChain and the AI Cup result. Some of it is academic work from TU Delft. Some of it is private software that can only be described through architecture. The public claim is simple: this is a lab for building the systems that let agents do real work.