Why Do Coding Agents Still Suck, and How to Improve Them

• ​Context engineering for environment-aware reasoning, bringing in richer information about system state, dependencies, and configurations to guide agent prompts.

• ​Tackling the limits of traditional RLVR in complex multi-step setups by exploring multi-turn RLVR approaches.

• ​Optimising thought-action-result trajectories to avoid common anti-patterns like repeated actions or premature termination.

• ​⁠What these advances mean for building coding agents that are actually useful in real-world workflows.

​About the Speaker:
Rohan Sharma (Student @ UCLA, Batch 6 Resident)
https://x.com/rs545837
https://www.linkedin.com/in/sharmarohan003

Pre-read material:
⁠https://arxiv.org/pdf/2503.15478
⁠⁠https://arxiv.org/abs/2506.11425
⁠⁠https://arxiv.org/abs/2507.09063