Let’s talk about efficient agent setups

1 points by Student

I’m fortunate that at work my llm spending budget is fully adequate to my needs. However, I’d like to go faster at work and I’d like to bring agentic coding into my own projects (frankly it’s going to be the only way I have time for side projects), so for that I want to be cheap and fast.

My recent setups:

  1. Cursor with composer 1.5 - very fast, capable, decent tool use, integrates its own code indexing features to make things faster. I’ve tried it with composer 2 but frankly it seemed slower and no better. Because my company’s litellm isn’t exposed to the internet, I haven’t tried it with open weight models, and using Claude with cursor blows my budget before the end of the month. Unfortunately I can’t see the details of the budget or my own individual spending. A memory hog.

  2. Recent Claude code with recent Claude sonnet and opus (4.6, 4.7). I’ve installed the various superpowers, and the only additional indexing/efficiency extension I’ve installed is the Serena MCP server. Serena is at least being used. Good for fairly big tasks with an instruction to tell it to plan, then review the plan then execute with subagents. Fairly expensive, costs about $600/month. What I’m currently using at work. Not super fast by any means. I relatively rarely find that it’s not able to do things I expect or getting stuck in loops. Code quality is decent.

  3. Older version of Claude with qwen series models accessed via litellm. No particular efficiency plugins. Fairly fast, fairly capable. Claude would often decide at random that there was an api error and not recover.

I’m interested in both your setups and how you’re measuring efficiency. Other than looking at per period spending I’m not measuring anything yet.