Entity: ARISE
ARISE is an AI engineering organization known for developer tool insights and agent evaluation frameworks, notably developing the AI agent Alex.
Agent Alex & The To-Do List Looping Failure Mode
During a multi-trace summarization run, ARISE’s agent Alex exhibited a severe agent failure mode: over the course of 27 model calls, almost all execution tokens were consumed reorganizing its internal to-do list rather than executing the actual summarization assignment.
The root cause was context crowding: as tool outputs, intermediate trace data, and past actions accumulated in the transcript, the original instructions and task priority were buried under noisy history.
Technical Solution: External Plan State Maintenance
To solve this looping behavior without overloading the prompt, the ARISE engineering team decoupled the active execution plan from the conversation transcript:
- Off-Transcript Disk Storage: The current plan was moved out of the conversation window and saved as a state file on disk (
current.md). - Fresh Plan Injection: On every model invocation, system middleware reconstructed a concise plan header directly from the disk state and appended it in front of the transcript history.
This architecture ensured that Agent Alex always evaluated its next move against the latest plan state rather than getting lost in transcript history, providing a primary technical case study for progressive-context-shaping.