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

Lecture 2: Agentic AI

From Chatbot to Co-Worker

Instructor: Ramdas Mozhikunnath
Course: Stage 1 — AI-Era Design Verification Foundations (DV-Gym)
Focus: Closed-loop agents, tool use, and human-in-the-loop review

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

Lecture 2: Agentic AI: From Chatbot to Co-Worker

💬 The Single-Shot Assistant

  • Mode: Passive Q&A (Prompt → Response).
  • Workflow: You copy code, run the simulator, paste the log back.
  • Bottleneck: Human engineer is the manual data bridge.

🤖 The Autonomous Agent

  • Mode: Closed-Loop (Goal → Plan → Act → Observe).
  • Workflow: Inspects files, runs scripts, analyzes simulator errors.
  • Enabler: Dynamic tool execution with human approval gates.
www.verificationexcellence.in
DV-Gym | Stage 1: AI-Era Design Verification Foundations

What Enables Autonomy: Tools & Memory

🛠️ Dynamic Tool Use

  • Moves beyond text generation to system action.
  • Inspects repo files, executes scripts, and queries databases.
  • Runs EDA simulator or linter commands directly in the sandbox.

🧠 Intermediate Memory

  • Retains observation history across multiple turns.
  • Remembers compile flags, prior errors, and tried paths.
  • Adjusts subsequent actions based on simulator feedback.
www.verificationexcellence.in
DV-Gym | Stage 1: AI-Era Design Verification Foundations

Under the Hood: How Agents Use Tools

  • LLMs Do Not Run Bash Directly: They cannot touch your shell natively.
  • Function Calling: Tools are exposed to the model as structured JSON schemas.
  • Structured Output: The LLM emits a tool call rather than plain conversational text:
    {"tool": "run_sim", "args": {"test": "axi_burst_test"}}
www.verificationexcellence.in
DV-Gym | Stage 1: AI-Era Design Verification Foundations

The Closed-Loop Execution Cycle: Flow

1️⃣ Emit & Execute

  • Agent Intent: Emits a structured command to compile the testbench (compile_tb).
  • Harness Action: Execution environment launches simulator (vcs / xcelium).
  • Status: Simulator runs in sandboxed terminal space.

2️⃣ Capture & Observe

  • Raw Stream: Harness captures stdout / stderr.
  • Diagnostic Signal: Isolates explicit failure line:
    Error: [SV-NF] Signal 'awlen' not declared
www.verificationexcellence.in
DV-Gym | Stage 1: AI-Era Design Verification Foundations

The Closed-Loop Execution Cycle: Convergence

3️⃣ Adapt & Patch

  • Hypothesis: Target variable is missing from interface scope.
  • Action: Agent edits tb_top.sv to declare logic [7:0] awlen.
  • Re-trigger: Automatically invokes recompile tool.

4️⃣ Gated Completion

  • Verify: Checks for clean exit code (0 errors, 0 warnings).
  • Halt Condition: Loop terminates autonomously when clean.
  • Handoff: Pauses execution and awaits human verification.
www.verificationexcellence.in
DV-Gym | Stage 1: AI-Era Design Verification Foundations

DV Pattern 1: Regression Log Triage

🎯 Scope & Tooling

  • Primary Goal: Group and root-cause nightly regression failures across 1,000+ seeds.
  • Required Tools: grep_log, view_trace, git_blame.
  • Input: Unstructured directory of raw simulation output logs.

🔄 Agent Workflow

  • Cluster: Groups logs by distinct error signatures (UVM_ERROR, timeouts, fatal assertions).
  • Correlate: Maps failure timestamps against recent RTL or TB git commits.
  • Deliverable: Ranked triage report with exact waveform markers to inspect.
www.verificationexcellence.in
DV-Gym | Stage 1: AI-Era Design Verification Foundations

DV Pattern 2: Testbench Scaffolding

🎯 Scope & Tooling

  • Primary Goal: Generate and wire a complete UVM interface agent from scratch.
  • Required Tools: read_spec, write_file, compile_sv.
  • Input: Interface pin list and transfer protocol timing notes.

🔄 Agent Workflow

  • Draft: Scaffolds driver, monitor, sequence items, and config classes.
  • Integrate: Hooks up the agent to the environment and virtual sequencer.
  • Self-Heal: Iteratively resolves missing semicolons and type mismatches.
www.verificationexcellence.in
DV-Gym | Stage 1: AI-Era Design Verification Foundations

DV Pattern 3: Coverage Gap Closure

🎯 Scope & Tooling

  • Primary Goal: Target and close lingering holes in functional coverage.
  • Required Tools: parse_ucdb, read_sequence, run_sim.
  • Input: UCDB / coverage database reports and existing sequences.

🔄 Agent Workflow

  • Analyze: Identifies uncovered cross bins in state machines and FIFOs.
  • Tune: Alters sequence constraint distributions, weights, and burst modes.
  • Validate: Re-simulates targeted seeds to confirm bin count increments.
www.verificationexcellence.in
DV-Gym | Stage 1: AI-Era Design Verification Foundations

DV Pattern 4: Spec-to-SVA Generation

🎯 Scope & Tooling

  • Primary Goal: Convert natural language protocol rules into formal/dynamic SVA.
  • Required Tools: read_spec, write_sva, run_formal.
  • Input: Bus protocol interface specification chapters and timing diagrams.

🔄 Agent Workflow

  • Translate: Generates concurrent properties (assert property) with clocking and reset.
  • Audit: Flags vague or contradictory timing requirements to the designer.
  • Sanity Check: Runs quick formal smoke runs to weed out vacuous passes.
www.verificationexcellence.in
DV-Gym | Stage 1: AI-Era Design Verification Foundations

Guardrails for Autonomous Agents

🔒 Execution Boundaries

  • Iteration Limits: Enforce hard stops (max 3–5 iterations) to prevent runaway billing and thrashing.
  • Sandboxed Tooling: Restrict tool commands strictly to safe verification directories—no system-level writes.

🛡️ Core Immutability

  • Protected Targets: Design RTL, golden scoreboards, and reference models are strictly marked Read-Only.
  • Mutant Testing: Run agent-modified tests against injected RTL mutants to verify checks are alive.
www.verificationexcellence.in
DV-Gym | Stage 1: AI-Era Design Verification Foundations

Human-in-the-Loop (HITL): The Golden Boundary

🤖 The Agent’s Role: Proposer

  • Handles repetitive boilerplate, initial drafting, and log triage.
  • Generates candidate testbench updates as isolated pull request diffs.
  • Operates strictly under defined tool permissions and observation budgets.

👤 The Engineer’s Role: Gatekeeper

  • Audits generated properties for vacuous pass loopholes.
  • Validates that test passes stem from genuine stimulus closure—not check deletion.
  • Retains exclusive sign-off authority for tapeout metrics and golden updates.

Core DV Principle: Autonomy accelerates the exploration; human verification guarantees the silicon.

www.verificationexcellence.in