Pragma removes the manual passing of context, not the choice of model

The most irritating part of working with several AI agents is not switching between models but having to explain to each of them all over again what has already been done. Pragma's author uses Codex, Gemini and the DeepSeek API and ran into exactly this: different CLIs, isolated memory and constant copying of context.

Pragma tries to be a layer on top of those tools. The user interacts with a single interface while the system itself invokes different harnesses and models, passes context to them and collects memory into shared storage.

That is an important shift in how the task is framed. If the tool works as intended, you will need to pick a model by hand less often: one agent can research a question, another write the code, a third check the result, without making the user a courier between them.

But shared context is not simply a folder where all the messages get dropped. You have to know which information counts as current, what to do about contradictions, and how much surplus context to send each model. Pragma's description so far offers no data on the quality of that synchronisation, on supported harnesses, on how memory is stored, on security or on behaviour during failures. The Hacker News discussion adds no further scrutiny either: the post has just two points and no comments.

So for now Pragma is interesting above all as an attempt at a real infrastructure problem. Its value will be determined not by the number of models connected but by how reliably the system keeps the thread of work running between them.

In your work, does more time go on choosing a suitable model or on carrying context between tools? Source: Hacker News - Newest: ""AI" "LLM""