Population
Long-running agent tasks involving planning, tool calls, interruptions, intermediate observations, and changing local subgoals.
All projects
seeking partner
Do long-running LLM agents still pursue the user's original goal after dozens of steps?
Measure goal drift across long-running LLM agent trajectories and test lightweight checkpoints that restore user intent without restarting the workflow.
HYPOTHESIS
Periodic intent checkpoints triggered by detected plan drift will reduce goal-inconsistent terminal actions by at least 25% with less than 5% additional model tokens.
PRIMARY METRIC
Rate of completed tasks whose terminal action remains consistent with the original user goal and constraints.
MEANINGFUL THRESHOLD
At least 25% fewer goal-inconsistent terminal actions with less than 5% additional tokens.
BUSINESS TARGET
More reliable long-running LLM agents and clearer evidence for when orchestration should pause, resume, or request human input.
DECISION RULES
Goal drift falls within token budget
Replicate across agent frameworks and longer trajectories.
Drift falls but checkpoint overhead is high
Trigger checkpoints only on higher-risk state changes.
No meaningful drift reduction
Do not add checkpoints as a general orchestration mechanism.
Business outcomes are research targets, not guarantees. A null or negative result may still create substantial value by preventing investment in an ineffective product, feature, or campaign.
Intervention
Compact intent checkpoints that compare the current plan and intended terminal action with the user's original constraints.
Comparator
The same long-running workflow without explicit goal checkpoints.
Secondary metrics
Task success · Recovery after interruption · Additional tokens · Checkpoint false-positive rate