DV-Gym | Stage 1: AI-Era Design Verification Foundations

Lecture 4: Orchestrating Agents & System Architecture

What Is Orchestration?

  • Multi-Agent Coordination: Organizing specialized agents into structured pipelines.
  • Context Isolation: Keeps individual prompts lean, focused, and free of extraneous scratchpad logs.
  • Handoff Governance: Manages input/output schema translation between tasks.

Why Avoid the "Mega-Agent"?

  • Context Pollution: Single prompts carrying planning, coding, and debugging drown in tokens.
  • Attention Degradation: Multi-task instructions lead to missed verification constraints.
  • Audit Deficit: Self-evaluating agents routinely overlook their own logical oversights.
www.verificationexcellence.in
DV-Gym | Stage 1: AI-Era Design Verification Foundations

Orchestration Flow: The DV Lead in Action

DV Lead orchestrator flow

  • Orchestrator splits the plan; specialists draft SVA and TB; harness runs sim; human signs off.

 

www.verificationexcellence.in
DV-Gym | Stage 1: AI-Era Design Verification Foundations

How They Work Together

Orchestrator specialist harness loop

  • Simple rule: Orchestrator divides and routes. Sub-agents draft. Harness executes safely. Engineer gates merge.

 

www.verificationexcellence.in
DV-Gym | Stage 1: AI-Era Design Verification Foundations

The DV Lead Mental Model: How Orchestration Works

The Orchestrator as the DV Lead

  • Task Decomposition: Takes a high-level goal ("Bring up PCIe Gen5 physical layer verification") and splits it into work packages.
  • Role Assignment: Assigns sub-tasks to specialized sub-agents based on their tooling and playbooks.
  • Dependency Management: Ensures the SVA agent doesn't run until the spec-parser agent extracts interface timing.
  • Gatekeeping: Reviews intermediate outputs before triggering compute-heavy simulator runs.

The Harness as Compute & Lab Infrastructure

  • Tool Sandbox: The safe shell environment where sub-agents execute vcs, xcelium, or Python scripts.
  • Resource Enforcement: Kills hung simulations, protects EDA license pools, and isolates scratch git worktrees.
  • Context Filtering: Slices raw megabyte-scale simulation logs so sub-agents only receive actionable failure windows.
  • Analogy: If the Orchestrator is the DV Lead, the Harness is the EDA compute farm & regression manager.

 

www.verificationexcellence.in
DV-Gym | Stage 1: AI-Era Design Verification Foundations

Core Orchestration Topologies

🧭 Router & Sequential Pipeline

  • Dynamic Router: Top-level classifier inspects failure signatures and delegates to the appropriate specialist.
  • Linear Assembly: Pipeline passes artifacts forward:

  • State Filtering: Summarizes previous steps instead of forward-propagating entire transcripts.

⚖️ Generator vs. Critic (Adversarial)

  • Generator Role: Drafts testbench sequences, drivers, and assertions from spec text.
  • Critic Role: An independent reviewer instructed to identify vacuous passes, dead code, and CDC bugs.
  • Core DV Rule: Never allow the agent that generated the code to be the sole auditor.
www.verificationexcellence.in
DV-Gym | Stage 1: AI-Era Design Verification Foundations

Multi-Agent DV Pattern: Plan → Build → Audit

Planner & Builder

  • Planner Agent:
    • Ingests protocol chapters and register maps.
    • Generates cross-coverage goals, corner cases, and SVA lists.
  • Builder Agent:
    • Implements sequence classes, transaction items, and assertions.
    • Compiles against simulator headers to clear syntax errors.

Auditor & Human Sign-off

  • Auditor Agent:
    • Runs ai-tb-audit skill to check for shallow copies and race hazards.
    • Verifies assertions contain proper disable-on-reset criteria.
  • Human Gatekeeper:
    • Audits the clean pull request diff.
    • Signs off on regression suite integration.
www.verificationexcellence.in
DV-Gym | Stage 1: AI-Era Design Verification Foundations

Multi-Agent DV Pattern: Automated Failure Triage

Stage 1: Classification & Routing

  • Triage Agent: Ingests regression log and seeds; isolates the primary error line.
  • Root-Cause Slicing:
    • If UVM_FATAL timeout Routes to TB-Debug Agent.
    • If Data Compare mismatch Routes to RTL-Debug Agent.
    • If Compile/License drop Routes to Infra Agent.

Stage 2: Synthesis & Reporting

  • Specialist Agent: Pulls waveforms or git history to locate the regression origin.
  • RCA Reporter Agent:
    • Compiles root cause, affected files, and recommended markers.
    • Posts structured ticket to bug tracking system.
  • Human Action: Verification engineer validates the diagnosis.
www.verificationexcellence.in
DV-Gym | Stage 1: AI-Era Design Verification Foundations

Stage 1 Architecture: How the Stack Fits (1/2)

Component Operational Role DV Workflow Mapping
LLM Next-token statistical prediction engine Drafts boilerplate SV/UVM code, summaries, and initial testplans.
Agent Closed-loop execution with tool capabilities Triages simulation errors, fixes syntax, and re-runs test suites.
Skill Packaged task-specific playbook Codifies senior DV wisdom (e.g., uvm-hang-triage, sva-from-spec).
Workflow Multi-step pipeline with validation gates Regulates step transitions ().
www.verificationexcellence.in
DV-Gym | Stage 1: AI-Era Design Verification Foundations

Stage 1 Architecture: How the Stack Fits (2/2)

Component Operational Role DV Workflow Mapping
MCP Universal tool & context interface standard Connects agents securely to VCS, Xcelium, Git, and Spec databases.
Orchestration Multi-agent collaboration & state routing Coordinates Generator vs. Critic setups and pipeline handoffs.
HITL Mandatory human gatekeeper authority Final sign-off on code merges, golden scoreboards, and tapeout metrics.

Bottom Line: AI accelerates verification tasks; the verification engineer retains ultimate sign-off authority.

 

www.verificationexcellence.in