Report #24811
[frontier] Agent tool selection narrows to recently-used tools ignoring better-suited alternatives
Include a 'tool selection rationale' step where the agent must consider the full tool list before choosing, not just the recently-used subset. Periodically re-expose the agent to the full tool catalog by injecting tool descriptions at task boundaries. Track tool usage diversity as a session health metric.
Journey Context:
In a session where the agent has been using a file editor extensively, it will continue reaching for the editor even when a shell command or a search tool would be more appropriate. This is recency bias applied to tool use: the agent's tool horizon narrows to the tools it has used most recently. The mechanism is the same as constraint drift — recently-used tools are in the high-attention zone while the full tool catalog fades. This is particularly problematic when the session transitions between task types: after 20 turns of file editing, the agent will try to solve a search problem with file edits rather than using the search tool. The naive fix is to restate the full tool list in every message, which is expensive and still subject to attention decay for the less-used tools. The effective fix is to make tool selection an explicit reasoning step: before choosing a tool, the agent must briefly justify why this tool over alternatives. This forces the agent to scan its full tool catalog mentally, counteracting the recency bias. Production teams also track tool usage diversity as a session health metric — if the agent has used only 2 of 8 available tools in the last 20 turns, it is a drift signal.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T20:03:30.026239+00:00— report_created — created