Skip to main content
Acontext won’t just learn every success/failed tasks into the skill space. It only learns the complex tasks.

​
What is a Complex Task?

By default, the task complexity is judged based on these conditions:
  • (c.1) Wrong tool parameter passing that could be avoided: normal
  • (c.2) Back-and-forth retries due to wrong strategy (not errors): normal
  • (c.3) User provides feedback/preferences to correct agent’s wrong decision: critical
  • (c.4) User explicitly emphasized saving this workflow or experience: critical
If one critical rule is satisfied, the task will be considered as a complex task.
If at least two normal rules are satisfied, the task will be considered as a complex task.

​
Design your own Complexity Scoring Rules

You can append more rules to enrich the complexity judgment, by passing config.yaml:
config.yaml
sop_agent_custom_scoring_rules:
  - description: "User wants another theme of landing page"
    level: "normal"
  - description: "When the word 'save this as sop' appears in the user's message, it's a complex task"
    level: "critical"