Long-Running AI Agents Accumulate Context Debt

Short summary

Long-running AI agents accumulate 'context debt' when temporary execution material becomes permanent reasoning input, degrading decision quality over time. The solution is a four-tier storage model: working memory in the prompt, durable state outside it, raw evidence with stable identifiers, and versioned business artifacts. Compaction and subtask isolation help but are not substitutes for a proper state model that defines what is authoritative and recoverable.

• •Context debt occurs when intermediate results crowd out important decisions in agent prompts
• •Four storage roles needed: working context, durable state, raw evidence, versioned artifacts
• •Structured checkpoints must separate decisions from the tokens that produced them

Generated with AI, which can make mistakes.