EDAgent

A deployable, skill-based research agent system for EDA workflows.

License

EDAgent is released under the MIT License. You may use, modify, and redistribute this repository under the terms in LICENSE.

Workflow Decision Logic

• The agent first decides whether the task matches a known workflow.
• If known, it directly selects the corresponding skill subset and executes.
• If unknown, it runs a temporary safe flow, captures feedback, and then creates a reusable new workflow + SOP.
• Future similar tasks are routed through this new known workflow.

Knowledge/Tool Guard Loop

• Before execution, the agent checks the knowledge base and tool registry first.
• It prefers reusing existing methods/tools to reduce hallucination and avoid duplicate implementation.
• After execution, it writes lessons back to KB/tools/SOP.
• SOP Hardened feeds into the next workflow step (PLAN), so later runs become smoother.

Why EDAgent

EDAgent turns a complex research workflow into a practical product-like experience:

• one-command bootstrap,
• natural-language interaction,
• structured execution with audit artifacts,
• continuous improvement from user feedback.

Persistent SOP Across Chats and Projects

One core advantage of EDAgent is SOP persistence beyond a single chat session.

• SOP is stored in repository assets, not only in conversation memory.
• Even if you switch chats or move to another project, workflow policy and operational habits are retained through repo-based governance.
• New experience is continuously hardened into workflows, skills, and SOP records, so the system does not reset to zero each time.

No Skill Selection Burden for Users

Users do not need to know, choose, or sequence skills manually.

• The agent decides the workflow class from your task.
• Then it selects the right skill subset and next step automatically.
• You can stay focused on goals and constraints in natural language.

30-Second Start

git clone https://github.com/Mr-Fang-VLSI/EDAgent.git
cd EDAgent
python3 run_demo.py

What run_demo.py does:

• verifies core folders,
• runs infrastructure checks,
• refreshes knowledge index,
• prints next-step guidance.

Core Capabilities

• End-to-end research orchestration: idea -> hypothesis -> experiment -> validation -> retro.
• Dynamic infrastructure maintenance: docs, knowledge, paper, and tool libraries.
• Governed execution: theory-veto gates, audit trails, and rollback-aware updates.
• User-facing reporting: targeted summaries and slide-ready outputs for specific questions.

Product Experience

• Conversational: users work in plain natural language.
• Adaptive: behavior evolves with user preferences and feedback.
• Self-maintaining: the system can refine SOPs and operational assets over time.

Typical Use Cases

• Build and iterate EDA research plans quickly.
• Keep paper/knowledge/tool artifacts organized and searchable.
• Run experiment loops with post-run reflection and next-step recommendations.
• Generate concise explanation decks for collaborators/advisors.

Workflow-First Orchestration

EDAgent routes in two phases for stability and scalability:

1. classify the task into a workflow class,
2. choose the next skill only from that workflow's skill subset.

This avoids global skill search on every turn and keeps routing predictable as skills grow.

If You Use Codex/Claude-Style Agents

After cloning the repo, you can start auto-deployment with one sentence in chat:

开始部署EDAgent

or

Start deploying EDAgent

Expected behavior after this trigger:

• verify repo/environment status,
• bootstrap/verify infra folders,
• run guard/audit/index checks,
• ask your research direction and hard constraints.

After clone + python3 run_demo.py, ask the agent to continue from your direction and constraints.

Example prompt:

My research direction is placement for dynamic-power reduction.
Constraints: area/timing must not regress.
Please start with a scoped plan, run the first validation loop, and summarize key findings.

Repository Layout

• AGENTS.md: top-level governance and orchestration policy.
• skills/: modular capabilities (execution, infra maintenance, domain methods).
• scripts/common/: reusable infra and indexing utilities.
• docs/knowledge_base/: protocol and landscape knowledge.
• docs/tool_registry/: tool metadata/catalog.
• slurm_logs/00_meta/: governance and audit artifacts.

Open-Source Scope

• Research and experimentation only.
• No production SLA or warranty.
• Validate outputs independently for high-stakes decisions.

See also:

• CONTRIBUTING.md
• ROADMAP.md
• docs/WORKFLOW_CATALOG.md