Skip to content

AgenticPD: A Stage-Aware Agentic Framework for Physical Design QoR Optimization

🕒 Published (v1): 2026-07-06 07:55 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

AgenticPD is a multi-agent framework that reformulates physical design (PD) QoR optimization as a structured, stage-aware search over a branching tree rather than flat parameter tuning. A Judge Agent selects which stage to branch from and dispatches hints to specialized Stage Agents (FP, PL, CTS, RT), which execute via stage-local skills; a Harness Skill manages observations, context, and checkpoint reuse. Across six benchmarks on ASAP7 and SKY130HD, AgenticPD achieves superior post-route timing compared to vanilla LLM baselines and prior SOTA tuners while remaining competitive on power and area.

Problem

Physical design QoR optimization spans four sequential stages (floorplan, placement, CTS, routing) with cross-stage dependencies: a locally optimal decision at one stage can degrade final post-route PPA. Existing methods either treat the flow as a flat black-box (AutoTuner, FastTuner) or as a script-generation task (ORFS-Agent, ChatEDA), neither exploiting stage-wise intermediate state nor enabling non-sequential exploration with checkpoint reuse.

Method

AgenticPD maintains an optimization tree \(\mathcal{T}\) rooted at the post-synthesis netlist, where each node \(n_s^k = (a_k(s), Q_k(s))\) stores the action and observed stage-level QoR tuple \(Q^{(s)} = (Q_1^{(s)}, \ldots, Q_m^{(s)})\) (WNS, TNS, power, area).

Judge Agent receives compressed history \(\mathcal{H}_k\) and the adaptive profile \(\mathcal{A}_k = \{E(n)\}_{n \in \mathcal{T}}, \{B(s)\}_{s \in \mathcal{S}}\), where \(E(n)\) tracks exploration balance per node and \(B(s)\) is the stage bottleneck (gap from best-known per-stage QoR). It outputs a decision \(\mathcal{D}_k = (\hat{n}_k, b_k, h_k)\): the node to branch from, the branching stage \(b_k\), and per-downstream-stage hints \(h_k\).

Branching: stages in \(\text{Bef}(b_k)\) are reused from the checkpoint at \(\hat{n}_k\) at zero cost; only \(\{b_k\} \cup \text{Aft}(b_k)\) are re-executed.

Stage Agents receive a fixed-size context \(ctx_s = (\{Q_k(i)\}_{i \in \text{Bef}(s)}, e_s, \text{hint}_s)\) (upstream QoR, cross-iteration experience, Judge's hint), decide action \(a_k(s) = \pi_s(ctx_s)\), and invoke the corresponding PD Skill to execute and retrieve \(Q_k(s)\).

Agent Harness provides two instruments: (1) an Observation Tool that assembles the tiered tree snapshot and computes \(\mathcal{A}_k\), controlling token growth by compressing less-visited nodes; (2) an Orchestrator Tool that pipelines Stage Agent calls after the Judge decides. Stage Agents use per-stage PD Skills (FP/PL/CTS/RT Skill) that define action spaces in IEEE CEDA METRICS2.1 format and wrap OpenROAD backend execution. The incumbent is updated by a timing-primary rule: advance when primary metric \(Q_1\) improves and guardrail metrics remain within tolerance (Pareto dominance as the special case).

Key Contributions

  • Formulation of PD QoR optimization as a structured multi-stage decision problem over a branching tree with reusable intermediate checkpoints.
  • A Judge Agent that navigates exploration/exploitation via adaptive signals (\(E(n)\), \(B(s)\)) rather than flat sampling.
  • A Harness Skill (Observation Tool + Orchestrator Tool) providing structured observations, compressed history, and context management with bounded per-agent token cost.
  • Four stage-local PD Skills that expose action spaces and backend execution interfaces per stage.
  • Demonstrated timing improvements over vanilla LLM baselines and prior SOTA tuners (AutoTuner, ORFS-Agent) on ASAP7 and SKY130HD across three designs.

Results

  • vs. Vanilla LLM (ASAP7, 20-iteration budget, 3 LLM backbones): AgenticPD outperforms in 7 of 9 pairs on timing; geomean improvement: Performance (ECP) +2.4%, Area +2.2%, Power +1.1%. Achieves positive post-route WNS (timing closure) on JPEG ASAP7 across all three backbones; Vanilla LLM fails or barely closes.
  • vs. AutoTuner and ORFS-Agent (Table 2, 6 benchmarks): AgenticPD achieves the best post-route WNS/TNS across all six benchmarks; on JPEG ASAP7, AgenticPD reaches WNS = +50.569 ps vs. ORFS-Agent's +0.337 ps and AutoTuner's −9.996 ps.
  • Convergence: All six benchmarks converge to within 20% of final PPA values by iteration 8–14 out of 20.

Limitations

  • Fixed 20-iteration budget; cost per iteration is high because every candidate is evaluated through full detailed routing signoff.
  • Evaluated only on three designs (AES, ibex, JPEG) on two tech nodes; generalization to more complex or proprietary PDKs is unverified.
  • LLM backbone sensitivity is observed (DeepSeek-V3.2 outperforms Qwen3.5 and Kimi-K2.5); framework performance is partly backbone-dependent.
  • The branching tree can grow large; tiered snapshot compression heuristics are described qualitatively but not formally analyzed for completeness.
  • No comparison against ML-based tuners beyond AutoTuner (e.g., Bayesian optimization baselines like REMOTune or PPATuner are cited but not benchmarked).

Relevance to Harnesses / Meta-Harnesses

AgenticPD is a direct instance of a domain-specific agent harness: the Harness Skill explicitly mediates between the Judge Agent and Stage Agents by managing observation construction, context assembly, orchestration of downstream agents, and token budget control — responsibilities that are harness-layer concerns rather than agent-logic concerns. The separation between the Orchestrator Tool (routing decisions to agents) and PD Skills (wrapping backend execution per stage) exemplifies the harness-as-infrastructure pattern, where the harness owns execution, checkpointing, and observation formatting while agents own reasoning. This design is directly analogous to meta-harness patterns that compose specialized sub-harnesses, making it a concrete EDA-domain case study for how to structure hierarchical agent scaffolds with bounded context growth and stateful branching.