ANNEAL: Neuro-Symbolic Agent Drops Recurring Failures to Zero with Governed Patches
ANNEAL converts recurring agent failures into typed, scored edits of a process knowledge graph with symbolic guardrails, canary testing, full provenance, and deterministic rollback. Tested across 4 domains and 27 runs, it is the only system in the study to commit persistent structural repairs. ReAct and Reflexion show 72–100% holdout failure on the same recurring faults; ANNEAL reduces this to 0%. Removing the governed-patch component loses up to 26.7 percentage points of task success.
Why It Matters
ANNEAL provides an architectural precedent for agents that learn from their own failure modes without requiring retraining — using a structured KG instead of uncontrolled prompt injection. The provenance and rollback mechanisms address the trust gap that makes self-modifying agents impractical in production today.