Struggling with Complex Instructions: Engineering Challenges in AI Agents
SemiMagnum · reddit · 2026-08-11
A developer shared on Reddit the bottlenecks encountered when trying to use AI (specifically Claude) to accelerate complex academic writing. The task requires 'merging' complex drafts with extremely tedious formatting guidelines.
- Agent Orchestration: Following Anthropic's docs, the author built a Claude Skill to break down tasks for sub-agents, but the model blatantly skipped basic instructions like creating project files at startup.
- Context Density: The author points out that the issue isn't the context window size, but the high 'density' of instructions and limitations, making it hard for the model to juggle all load-bearing rules simultaneously.
- Tool Comparison: The author also reviewed ChatGPT (heavy hallucination), Gemini (introducing new bugs while fixing old ones), and Claude (better but still not fully trustworthy), debating whether to fall back to 'deviation analysis only' or continue optimizing the agentic workflow.
More from coding & agent
- Scholar Offers $100 Reward as Claude Code & Codex Fail to Fix Mac Wi-Fi Drop — paulnovosad · 2026-08-11
- Codag: Open-Source Tool for Visualizing AI Workflows in VSCode — tom_doerr · 2026-08-11
- Lack of Agent Identity on GitHub Breaks Native Code Review Workflows — Vjeux · 2026-08-11
- mitsuhiko Shares Second Iteration of Earendil Durable Execution Framework — mitsuhiko · 2026-08-11
- LangChain Unveils Deepagents: The Second Cheapest Agent Harness — hwchase17 · 2026-08-11
- PM's Guide: Mastering Frontend State Management for AI Coding Agents — brandon_galang · 2026-08-11