Conversation

Jarkko Sakkinen

Edited 9 days ago

What they call agents, are actually markdown documents dumped at the beginning of every context (🤦‍♀️), or at most gradually pulled from a skill set on demand.

And it’s a very disturbing topic overall. I mean all this role playing of imaginary characters as part of daily work life is just… I don’ t know…

For stuff that I do with AI, which right now is to create an O with local LLM focus, Pi saves me from all that bullshit.

My Pi setup has neither subagents nor MCP servers. They are both useless and impose a fixed cost for every new context.

When I initialize a context I might dump relevant data instead of descriptions of agents and their personas (nnnggh…. thinking of this feels as physical pain tbh).

And this is how task management works:

  1. /fork: creates a new session.
  2. /tree: lets you time-travel through the context history and create new branches into the past to continue from.
  3. /merge: combines the running context with a session given as a parameter.
  4. Background task creation: I just prompt to PI “spawn a new process with fresh a session and …” or something.

The 3rd feature is /merge, which comes from the extension npm:pi-session-merge.

1
0
2

Jarkko Sakkinen

Edited 9 days ago
It's all I have for "subagents". E.g., I don't have agentic teams building anything at all in my life. If I use AI, I'm doing something proactively, and it's what it is i.e., a tool.

Thus, a small and focused set of raw as possible primitives provide a tactile controls, which is a priority, as such random need comes always unplanned, as I don't plan that much in my life :-) I just like to do stuff and be focused on target.
1
1
1
In place of agents, teams and complex software proceses I just initialize model of suitable complexity and copy-paste relevant contextual information (for the task). Then it does its thing and dies :-) I.e. I ignore process feelings of itself as an individual and focus on sound execution process...
0
0
1