The Failure Mechanism

• Long-running agents often fall into a loop of "runaway continuation" where they consume resources re-organizing their own to-do lists rather than completing the task, a phenomenon observed in the ARISE system where 27 consecutive model calls failed to progress the project.7:06
• The most common failure mode is a "graveyard of stale rules" where an initial prompt or history log from hours ago begins to outrank new evidence, causing the agent to pursue obsolete goals.1:44

Progressive Context Shaping

• To solve for this, successful teams at organizations like OpenAI and Anthropic split context into four distinct layers: stable instructions, current project state, a resource map, and historical decision logs.13:08
• Current state serves as the governing document; it is kept external to the transcript and rebuilt on every model call, ensuring the agent only acts on the latest definition of the goal and its constraints.7:44
• The human operator holds the governing judgment, deciding when new evidence necessitates a state update, while the agent acts as the execution engine for the paths between those checkpoints.16:20

Operational Limits

• Engineering teams often struggle to manage more than three to five concurrent agent sessions before context switching becomes cognitively painful, suggesting that parallel agent workflows require externalized coordination systems like ticket boards.17:19
• In one OpenAI project involving over 1,000,000 lines of code, replacing a massive manual with a short, living map and linked project documents enabled three engineers to ship in one-tenth of the time a manual approach would have required.2:42