Started 1w ago 4 replies 690 views
I'm making a small detective game and I want an agent that tests dialogue trees: it walks through conversations, flags dead ends and checks that clues appear before they're needed. The script is unreleased and I'd rather it never leaves my machine.
I've got a decent desktop GPU. Is a small open-weights model running locally good enough for multi-step agent work, or will I spend a month fighting it? Has anyone actually done this?
Dev RamanStaff · Editor
· Senior Editor, Agents
I spent a week offline with a laptop-sized open model (Field Notes here). Summary: great for small, well-defined jobs, much weaker on long chains of reasoning. Your dialogue-walking task can be split into small steps, which plays to its strengths.
My pattern: local model for the many cheap, repetitive checks (does this node have an exit? is the clue mentioned earlier?), and only escalate the genuinely hard judgement calls. For your case, the hard calls could just go to you.
Privacy is a solid reason to go local, but local isn't automatically safe. Keep the agent's file access scoped to the game's dialogue folder. A confused agent with write access to your whole project is a bad afternoon regardless of where the model runs.
Plan: local model, one check per step, results written to a plain text log, read-only access to the dialogue folder. I'll report back in Show & Tell if it finds the dead end I already know about.
