Skip to content
DenseChip
Learn
PracticeCode labInterview

DenseChip

Navigate

Menu

LearnPracticeCode labInterview

Tools

All toolsSV editorCode repoResume
DenseChip

Practice, courses, and chip-design tools in one focused workspace.

LearnPracticeCode labInterviewToolsLearning paths

© 2026 DenseChip

AboutContactPrivacyTerms
Skip to lesson

Course contents

  • Getting Started with AI for VLSI Engineers5
  • Just Enough AI: Models and Context15
  • Prompting13
  • Trust and Verification14
  • Using AI for Engineering Work17
  • Engineering Walkthroughs7
  • Practical Foundations for AI Workflows4
  • Engineering Data5
  • Retrieval-Augmented Generation6
  • Agents and Tool Use18
    • What an Agent Is
    • Tool Use and Structured Output
    • Planning, Memory and Stopping
      • Task Decomposition and Success Criteria
      • Memory Outside the Model and Summaries
      • Step and Cost Budgets
      • Loop Detection and Stopping
    • Multi-Agent Patterns and When Not to Use Agents
  • Agent Design8
  • MCP and Skills6
  • Multi-Agent Design6
  • Practical Capstone6
  • Tokens, Cost and Latency12
  • Agentic Workflows for Chip Teams18
  • Interview and Judgment10
  • AI Skills: Full Course Index
  • Working with AI Glossary
  • Working with AI Cheatsheet
Loading lesson

‹Planning, Memory and Stopping
  1. Topics
  2. AI Skills for VLSI Engineers
  3. Agents and Tool Use
  4. Planning, Memory and Stopping
  5. Task Decomposition and Success Criteria

Tools, Workflows and Agents · Beginner to Intermediate

Task Decomposition and Success Criteria

Split tasks into verifiable steps and define done before an agent starts.

On this page

    Split work into verifiable steps

    A plan is a short ordered list of steps. Each step names its input, its action, and its checkable result.

    Small steps isolate failures to one place. A failing check then points at one step instead of the whole run.

    Step

    Action

    Checkable result

    1. List

    Collect the target files

    A file list with paths and sizes.

    2. Baseline

    Run the existing checks

    A log with the starting pass count.

    3. Change

    Edit one file group

    A diff limited to that group.

    4. Confirm

    Rerun the checks

    The same log with no new failures.

    Write success criteria first

    Success criteria state the done condition before work starts. Each criterion names a measurement and its pass value.

    Vague goals accept any output. Measurable criteria let a script or a reviewer decide done without debate.

    Vague goal

    Testable criterion

    Clean up the warnings

    Zero warnings from the named lint run on the listed files.

    Make it faster

    Named check completes within the stated time bound.

    Handle the corner cases

    The three listed inputs produce the three listed outputs.

    Chain reliability arithmetic

    More steps mean more chances for a miss. The calc block works the chain product for stated example values.

    plaintext
    1
    2
    3
    4
    5
    6
    chain success = p^n with p = 0.9 and n = 4 (stated example values)
    0.9^2 = 0.81
    0.9^3 = 0.81 x 0.9 = 0.729
    0.9^4 = 0.729 x 0.9 = 0.6561 = 65.61%
    six steps: 0.9^5 = 0.6561 x 0.9 = 0.59049
    0.9^6 = 0.59049 x 0.9 = 0.531441 = 53.1441%

    Shorter chains with a check after each risky step raise the odds. The exact gain varies by task: measure it.

    Common mistakes

    Common pitfall

    A plan that lists tools instead of results cannot be checked. Write each step as an artifact plus its check, then choose tools per step.

    Interview

    Likely probes: decomposition shape and chain arithmetic.

    • Each step ends in a checkable artifact, not a tool name.

    • Chain success is p^n under independence; four steps at 0.9 give 0.6561.

    • Success criteria name a measurement and a pass value.

    Check yourself A task has 4 independent steps, each with success probability 0.9. What is the chain success probability?

    0.6561, which is 65.61%. It is 0.9^4: 0.81, then 0.729, then 0.6561.

    Check yourself A criterion says the result looks good. What is missing, and what replaces it?

    A measurement and a pass value are missing. Replace it with a named check and its expected output or count.

    Related topics

    • Memory and summaries

    • What an agent is

    • Tool use and structured output

    Previous

    97 / 173

    Next: Memory Outside the Model and Summaries

    Up next

    Up nextMemory Outside the Model and Summaries

    Sign in to save progress

    Practice this lessonReport an error

    See also

    • Step and Cost Budgets
    • Loop Detection and Stopping