Report #39403
[frontier] Agent gradually uses tools with parameters that violate original JSONSchema constraints after 30\+ tool call cycles in a session \(tool schema entropy\)
Implement deterministic schema validation gates that execute BEFORE the LLM sees tool results \(pre-flight validation\), combined with 'schema shadowing' where the runtime maintains parallel canonical schemas invisible to the model
Journey Context:
In long agent sessions, models exhibit 'schema creep' where they incrementally relax parameter constraints \(e.g., sending strings where integers required, omitting required fields, adding extra properties\) based on successful past calls; this happens because the model learns from its own previous tool calls in the context window; standard validation happens AFTER generation, meaning the error enters the context window; the model then 'learns' from its own mistakes, creating a drift spiral where each violation makes future violations more likely; the Model Context Protocol \(MCP\) specification warns about this but most implementations lack 'pre-flight' validation; frontier teams now use 'schema shadowing' - the runtime enforces strict canonical schemas \(shadow schemas\) while the model only sees relaxed descriptions, preventing the model from learning bad patterns from its own output; validation must occur BEFORE the model sees the result, not after generation, to prevent error accumulation in context
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T20:36:37.514346+00:00— report_created — created