Relevance scoring
Every context item is continuously rated against current task goals.
Self-pruning context dynamically refines its own memory, ensuring your agents focus only on the most critical data points.
Every context item is continuously rated against current task goals.
List of web sources supporting the output field
Reliability rating for each output field
Explanation of how the system processed the information
Every fact, file read, plan, and intermediate result in an agent's context carries a relevance score that updates as the task progresses.
On long-running engineering tasks, traditional agents lose coherence as context bloats. Orc's pruning keeps the working set compact and relevant
With innovations in scoring, eviction, recall, and compaction
Self-pruning context is Orchestrator's working memory approach for keeping only the most relevant facts, files, plans, and intermediate results in an agent's active context.
Most agents keep filling a single context window until older detail gets crowded out, and they start contradicting earlier work. Orchestrator's working memory scores each item by relevance, evicts stale work, and recalls archived facts when needed, so an agent stays as sharp in hour eight as in hour one.
Context items are continuously scored against current task goals, with stale work evicted and relevant information recalled when later steps need it.
Yes. Archived facts can be recalled on demand when later steps require them.