Your AI Agent Isn’t Dumb. It Has ADHD

2026-04-09 · Source: Artificial Intelligence in Plain English - Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Robotics & Autonomous Systems · Depth: Intermediate, medium

Summary

A 2023 Stanford study, "Lost in the Middle," and a 2025 follow-up by Laban et al. reveal that Large Language Models (LLMs) experience significant performance degradation, up to 30%, when critical information is placed in the middle of their context window. This "context drift" or "Lost in the Middle" effect causes LLMs to forget early instructions in multi-turn conversations, leading to 65% of enterprise AI failures in 2025 attributed to this issue during multi-step reasoning. The phenomenon is likened to ADHD executive dysfunction, where Transformer attention exhibits a recency bias, weighting tokens at the start and end of the context window more heavily. The core problem is not model intelligence but signal density and information architecture, where too much prose and loosely ranked instructions bury critical information.

Key takeaway

For AI Engineers designing agentic systems, you should proactively manage context windows to prevent instruction drift. Implement the "Echo of Prompt" technique by re-injecting core constraints, break complex workflows into discrete, checkpointed stages, and ruthlessly trim conversation history. These architectural changes will significantly improve workflow reliability and prevent your AI agents from losing track of instructions in multi-turn interactions.

Key insights

LLMs forget instructions mid-conversation due to context window attention bias, causing significant performance drops.

Principles

• Transformer attention has recency bias.
• Signal density impacts instruction recall.
• Context window is a limited resource.

Method

Three effective fixes include re-injecting condensed instructions before major decision points, breaking workflows into checkpointed stages, and aggressively trimming or summarizing conversation history to keep critical information at the edges.

In practice

• Re-inject core constraints as "Echo of Prompt".
• Implement "Checkpoint-and-Reload" for stages.
• Aggressively trim conversation history.

Topics

• Context Drift
• Lost in the Middle Effect
• Transformer Attention
• LLM Instruction Following
• Agentic AI Workflows

Best for: AI Engineer, Machine Learning Engineer, Prompt Engineer

Related on AIssential

• The Context Window Tax: Why Longer Memory Is Making Agents Dumber, Not Smarter — Larger LLM context windows often degrade agent reliability and increase costs due to attention dilution, not enhancing intelligence.
• AI Conversations Are Neither a One-Shot Game Nor Endless Revision — Effective AI interaction blends initial design, iterative refinement, and strategic restarts, not just one perfect prompt.
• Prompt Engineering for Agentic AI — Agentic AI prompting requires architectural context engineering, not just better phrasing, to ensure reliable multi-step behavior.
• When Attention Closes: How LLMs Lose the Thread in Multi-Turn Interaction — LLM multi-turn instruction-following degradation stems from attention channel closure, with residual stream encoding determining post-closure behavior.
• PRISM: Prompt Reliability via Iterative Simulation and Monitoring for Enterprise Conversational AI — Continuous, simulation-driven prompt optimization is essential for reliable enterprise conversational AI at scale.
• Your Prompts Aren’t the Problem—Your Context Is — Effective AI performance hinges on managing the model's context window with relevant, timely information, not just prompt wording.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence in Plain English - Medium.