CS2680 Student Blog
Student blog posts on optimizing LLM agent systems.
Topics
Assignment 1 · Mads-Lens
Web interfaces for your agents
Building a live web UI driven by Claude Code to visualize real-time tool trajectories, subagent branching, and session costs from the user’s perspective.
4 posts
Assignment 3 · Mads-Opt
High-efficiency agent orchestration
Engineering context compression, tool concurrency, and dynamic model routing around a black-box LLM API to slash latency and cost while preserving benchmark task accuracy.
0 posts
Assignment 5 · Mads-Stack
Optimizing the full agent stack
Profiling and optimizing the complete agent stack, from control flow and serving infrastructure down to open-weight inference,to maximize end-to-end compute and token efficiency.
0 posts
Latest posts
Or browse by author or by tag.
Controller: a Teenage Engineering-inspired terminal for Claude Code
Controller turns Claude Code's settings into hardware: an effort knob, model keys and a subagent fader set up each run, with cost, model and agents always in view, so you steer the agent with controls instead of prompts.
Why Claude Code Kept Reaching for Bash
If your Assignment 1 viewer filled up with Bash cards, your prompts were mostly not to blame. Claude Code told the agent to use the shell, dropped the shell tool's usual warning, and in most installs had no Grep or Glob. Here is what the session logs show.
Lessons for Agentic Coding from Student Projects
Lessons from CS2680 student projects: make decisions and context editable, show costs and delegation, define reliable event states, and verify the behavior users actually depend on.
Six interfaces for understanding coding agents
Six Assignment 1 projects explore how to make delegation, time, and control visible when working with a coding agent.Publishing your post
See publishing guide for the instructions.