The 'Trajectory' Effect: How AI Coding Agents Accumulate Bad Habits
mattpocockuk · x · 2026-08-06
Dex Horthy recently discussed the concept of an AI coding agent's "trajectory": agents tend to pick up habits during a session, such as repeatedly using cURL for verification if done once.
- Context Pollution: Clearing the context is sometimes necessary to change these trajectories. For instance, a personal podcast generator with access to past transcripts might pick up annoying catchphrases or over-focus on meaningless metrics due to accumulated state.
- Codebase Degradation: This effect is particularly dangerous for codebases. Codebases are stateful and accumulate state over sessions, meaning most codebases trend downwards on their trajectory.
- Solutions: The only way to fix a bad trajectory is through better steering and by changing the environment the agent operates in.
Related event: AI Coding Agents Suffer from Trajectory Effect and Context Pollution(2 posts)→
More from coding & agent
- Dev Advocates Ditching Claude for GPT or Kimi in Coding — MarcJSchmidt · 2026-08-06
- AI Toolkit Helper Released: Utility Tools for Model Trainer — ostrisai · 2026-08-06
- DeepSeek API Adds Responses Format with Built-in Web Search and Codex Support — teortaxesTex · 2026-08-06
- Shopify's Continual Learning Flywheel Beats Frontier Models, Cuts Costs 96% — MParakhin · 2026-08-06
- Don't Use Prompts to Govern Agents: Enforcement Belongs at the Tool Boundary — No-Conflict4823 · 2026-08-06
- Winning with Claude Code Subagents: Cap Scope, Don't Just Unleash a Swarm — PrajwalTomar_ · 2026-08-06