vipin wiki

Analog Agent

4 min read

Analog Agent

Summary

analog-agent is Vipin’s AI4EDA layered analog circuit design harness. It treats natural-language specification understanding, strict task compilation, world-model support, planning/search, simulator-backed verification, and memory/reflection as separate system roles rather than a single generative loop.

Current Contribution

  • EXTRACTED: The intended paper-facing claim is a layered analog-agent system grounded in real SPICE verification, with a world-model-guided and planner-mediated loop rather than a one-shot generator.
  • EXTRACTED: The safe default claim boundary is a calibrated, surrogate-guided analog sizing loop under explicit SPICE truth boundaries.
  • EXTRACTED: The six-layer spine is specification understanding, task formalization, world-model services, planning/optimization, simulation/verification, and memory/reflection.
  • EXTRACTED: Frozen runnable vertical slices include ota2_v1, folded_cascode_v1, ldo_v1, and bandgap_v1, with ota2_v1 retained as the primary submission-facing path.
  • EXTRACTED: The default comparison stack includes full_simulation_baseline, random_search_baseline, bayesopt_baseline, cmaes_baseline, rl_baseline, no_world_model_baseline, and full_system.
  • INFERRED: Compared with the LLM4Rec projects, analog-agent shares closed-loop decision and uncertainty discipline but has a physical SPICE/configured-truth evidence gate instead of a same-candidate ranking gate.

Current Status And Gates

  • EXTRACTED: The project is no longer only a front-end scaffold; the first six system layers are present in formal schema-and-service form with API routes, tests, vertical slices, and submission-package utilities.
  • EXTRACTED: The README reports the local-only archive/research/papers/submission_package/ demonstrator-truth package as external_submission_ready = true for the frozen ngspice path; treat this as target-project status, not as wiki-validated configured-truth or signoff readiness.
  • EXTRACTED: Stronger configured-truth / external-PDK / Spectre-oriented claims remain separate upgrades, not current default paper claims.
  • EXTRACTED: The repository intentionally does not create or download user-managed PDK roots, external model cards, Docker exposure, or local mount paths automatically.
  • EXTRACTED: If the repo reports demonstrator_only, configured_truth_not_ready, or configured_truth_candidate_ready, stronger configured-truth claims must remain withheld until assets are present and validated.

Canonical Startup Packet

Read before nontrivial work:

  1. README.md
  2. docs/configured_truth_user_action_boundary.md
  3. docs/repo-map.md
  4. docs/related_work_map.md
  5. docs/stop_conditions.md
  6. configs/default.yaml
  7. configs/benchmarks/multi_task_suite_v1.yaml
  8. configs/simulator/ngspice.yaml
  9. scripts/run_system_closure_report.py

For closeout or paper-readiness decisions, also run or inspect:

  • scripts/review_harness_closeout.py
  • scripts/check_configured_truth_readiness.py
  • scripts/check_open_pdk_ready.py
  • scripts/check_native_artifact_rerunability.py
  • scripts/check_container_runtime.py

Module Map

AreaRole
apps/api_server/FastAPI routes for interaction, tasking, world modeling, planning, simulation, memory, experiments, and acceptance checks.
apps/orchestrator/Cross-layer job and truth-loop orchestration.
apps/worker_llm/, apps/worker_simulator/, apps/worker_world_model/, apps/worker_memory/Thin worker-facing adapters; shared behavior should stay in libs/.
libs/schema/Typed contracts for specs, tasks, actions, world models, planning, simulation, memory, benchmark evidence, and submission packages.
libs/interaction/, libs/tasking/, libs/world_model/, libs/planner/, libs/simulation/, libs/memory/Core six-layer implementation spine.
libs/eval/, libs/vertical_slices/Benchmark runners, statistics, baselines, evidence packaging, and frozen OTA/folded-cascode/LDO/bandgap slices.
configs/Benchmarks, simulator truth boundary, PDK, LLM roles, world-model settings, and runtime defaults.
templates/Frozen netlist/testbench/measurement contracts for canonical verification paths.
scripts/Test suite, benchmark, evidence generation, export, readiness, closeout, and review-gate commands.
archive/, .artifacts/Ignored local outputs and transient simulator/cache artifacts. Inventory only from the wiki.

File Inventory Baseline

2026-05-17 live scan:

MeasureCount / summary
git ls-files376 tracked paths
rg --files358 visible searchable paths
tracked text-like paths359
tracked non-text/artifact marker paths17, mostly .gitkeep placeholders
category countssource 184; test 82; script 58; config 33; doc 7; template 4; infra 4; other 4
largest local artifactsignored archive/legacy/... and .artifacts/simulation/.../tran.json files around 0.2 MB each; no giant raw/model artifact was visible in this scan

File Area Rules

  • Safe to edit when working inside analog-agent: libs/, thin apps/ wrappers, configs, templates, tests, scripts, and canonical docs for the requested subsystem.
  • Caution: paper/export helpers and closeout scripts affect claim wording and reportability; inspect the relevant review tests before changing them.
  • Inventory only: archive/, .artifacts/, local .venv, local PDK/model-card paths, simulator logs, and generated benchmark/submission outputs.
  • Do not copy into public wiki: raw simulator logs, private paper drafts, local PDK/model-card contents, .env, generated archive outputs, or any proprietary/process-sensitive model data.

Current Git Reminder

2026-05-17 status:

## main...origin/main

Remote:

https://github.com/appleweiping/analog-agent.git

The worktree was clean during this wiki audit.

Future Agent Entry

Recommended local commands from current docs:

py -3.12 scripts\run_test_suite.py
.\scripts\run_test_suite.ps1 -UseVenv -RequireApiDeps
py -3.12 scripts\review_harness_closeout.py
py -3.12 scripts\check_configured_truth_readiness.py
py -3.12 scripts\run_system_closure_report.py

Acceptance and benchmark entrypoints:

py -3.12 scripts\run_ota_acceptance.py
py -3.12 scripts\run_folded_cascode_acceptance.py
py -3.12 scripts\run_ldo_acceptance.py
py -3.12 scripts\run_bandgap_acceptance.py
py -3.12 scripts\run_benchmark.py

Difference From Sibling Projects

  • uncertainty, truce-rec, and tgl-rec are LLM4Rec projects governed by same-candidate evaluation and baseline gates.
  • analog-agent is an AI4EDA/SPICE harness governed by physical-validity and configured-truth gates.
  • It is related to protein-optimization-feedback-shift through closed-loop uncertainty-aware optimization, but its evidence is simulator-backed analog circuit closure rather than biological sequence feedback.

Counterpoints and Gaps

  • The public repository intentionally abstracts some research-sensitive backend details; do not infer missing internals from README absence alone.
  • Demonstrator-truth external_submission_ready = true is not the same as configured-truth, external-PDK, Spectre, signoff, layout, PEX, or yield readiness.
  • This wiki pass did not run the test suite or inspect generated archive outputs; it preserved metadata-level routing and artifact boundaries.

Graph View

Backlinks