A foundational framework for converting natural-language intent into verifiable RTL and physical implementation via structural synthesis loops.
AgentIC structurally mitigates two primary failure modes inherent in contemporary LLM-driven EDA workflows.
LLM pipelines frequently yield false positives against weak assertions. AgentIC enforces deterministic semantic rigor, multi-corner Static Timing Analysis, and Logical Equivalence Checking as non-negotiable exit criteria.
Unbounded generative loops reliably diverge. AgentIC implements stateful failure fingerprinting and probabilistic decay boundaries to assert a definitive fail-closed state upon stagnation.
Pre-synthesis structural verification; enforcing exact width resolution and port topological integrity.
Deterministic limits on generation permutations derived from historical trajectory analysis.
Condensation of verbose EDA execution traces into actionable causal vectors.
Closed-loop parameter adjustment informed by localized congestion and WNS deltas.
Aggregated validation across power (IR drop), timing (Hold/Setup), and functional equivalence.
Automated modular partitioning dictated by synthesized module interconnect topologies.
AgentIC abstains from linear prompt chains. It orchestrates a Directed Acyclic Graph (DAG) with explicit fallback conditions and constraint mutation.
Disambiguates between stimulus faults (Testbench) and logical synthesis faults (RTL), generating discrete summary matrices for contextual repair.
Detects timing stagnation (ΔWNS < 0.01ns/epoch). Upon detection, induces structural modifications such as pipeline insertion or target density relaxation.