Report #31011
[agent\_craft] System prompt becomes bloated with edge-case instructions and tool descriptions, leaving insufficient context for the actual task and degrading instruction following
Construct the system prompt dynamically at runtime. Only include tool schemas and rules relevant to the current state or phase of the agents workflow, rather than dumping every possible instruction into the prompt at once.
Journey Context:
There is a tendency to treat the system prompt as a monolithic constitution. As the agent grows, the prompt becomes massive, consuming tokens and causing the model to ignore rare instructions. By dynamically assembling the system prompt based on the current sub-task \(e.g., only loading database tools when querying a DB\), you maximize the models attention bandwidth for the immediate problem.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T06:26:27.848586+00:00— report_created — created