AI in Transportation Lab — Texas State University

Agentic AI
in Civil Engineering

From autonomous structural analysis to intelligent construction management — exploring how LLM-powered agents are transforming every facet of civil engineering.

Amir Rafe
Postdoctoral Researcher
Spring 2026
Begin Exploration About me
Scroll to explore
Module 01

The Agentic AI Era

Autonomous systems that perceive their environment, reason about goals, take actions, and learn from outcomes — moving far beyond simple prompt-response AI.

From symbolic AI to long-horizon agent systems

Timeline of modern AI agents

Era 01

1950s-1980s

Symbolic and rational agents

Early AI framed intelligent systems as agents that perceive, maintain state, pursue goals, and choose actions through explicit reasoning.

What changed

This era was dominated by search, planning, knowledge representation, and expert systems built from hand-authored rules.

Representative ideas

Search, planning, rule engines, expert systems

Why it matters now

Modern agent design still inherits this core idea: intelligence is not just answering, it is acting under constraints.

Anthropic taxonomy OpenAI guide Transformer substrate ReAct Toolformer Reflexion Karpathy LLM course
Generated visualization of an agentic reasoning core coordinating tools, memory, and domain state

Generated system visual

One reasoning core coordinating context, tools, memory, and execution state

Agent runtime

Reliable taxonomy

From augmented LLMs to multi-agent systems

Augmented LLMs

The simplest useful layer is a language model connected to retrieval, tools, structured outputs, or shared state. This is broader than chat, but not always a full agent yet.

Workflows

When the sequence is mostly known ahead of time, routing, chaining, retries, and approval gates should stay explicit. Many so-called agents are better understood as workflows.

Single-Agent Loops

A true agent repeatedly plans, acts, observes, and decides when the task is complete. This is often the right default before introducing multiple specialist agents.

Multi-Agent Systems

Multiple agents become useful when the prompt logic, tool surface, or review burden becomes too large for one runtime loop. Typical patterns are manager-worker, handoff, and critique-reviewer systems.

Grounding, Memory, and Evals

Reliable agents stay grounded with retrieval and tool observations, stay coherent with summaries and memory, and stay safe through permissions, evals, and clear stopping conditions.

Typical agent system loop

Click a stage to inspect how a modern agent actually runs.

384+
Papers on arXiv
90%+
Agent Accuracy on SAP2000
10×
Faster FEM Setup
2026
Year of the Agent
Module 02

AI Agent Domains in Civil Engineering

Nine civil-engineering subfields where agent systems are already becoming useful, from structural sensing and digital twins to transport, construction, and evacuation safety.

Drag or use the arrow to explore →→→
01

Structural Health Monitoring

Agents fuse sensor streams, imagery, and maintenance history to flag anomalies, rank likely damage mechanisms, and trigger follow-up inspections across bridges, buildings, and dams.

Structural health monitoring scene with bridge sensors and monitoring overlays
SHM
02

BIM & Digital Twins

LLM-driven agents query BIM and IFC models in natural language, surface coordination issues, and keep digital twins current with field observations and operational data.

Building digital twin illustration with BIM overlays
BIM / DIGITAL TWIN
03

Autonomous FEM Analysis

Agentic simulation workflows can translate problem statements into boundary conditions, meshing, solver runs, and result summaries for faster structural analysis loops.

Finite element simulation illustration with structural stress heatmap
FEM / SIMULATION
04

Drawing & Design Automation

Design agents parse CAD files and markups, apply revision rules, and generate cleaner drawing updates for detailing, code checks, and reinforcement layout changes.

Engineering drawing automation scene with bridge plans and digital markup
CAD / DESIGN
05

Infrastructure Inspection

Vision-language agents review drone, mobile, and field captures to detect cracks, corrosion, and spalling, then prioritize defects for engineers and asset owners.

Drone-based infrastructure inspection scene focused on cracked concrete
INSPECTION
06

Traffic & Transportation

Multi-agent control systems optimize intersections, corridor flows, and transit coordination by combining simulation, sensing, and adaptive signal decisions.

Smart transportation scene with connected intersection and traffic overlays
TRAFFIC / TRANSPORT
07

Construction Management

Project agents summarize RFIs, progress, weather, and supply signals to highlight risk, recommend schedule moves, and support safer, more coordinated delivery.

Construction management scene with site logistics and schedule overlays
CONSTRUCTION
08

Geotechnical & Environmental

Agents interpret borings, maps, groundwater, and monitoring data to support slope stability reviews, site characterization, and environmental risk assessment.

Geotechnical engineering illustration with soil layers and foundation analysis
GEOTECH / ENV
09

Evacuation & Safety

Safety agents combine occupancy data, routing logic, and scenario reasoning to support evacuation planning, inclusive wayfinding, and emergency decision support.

Evacuation safety scene with indoor routing and occupant flow overlays
SAFETY / EVACUATION
Previous
More domains
Module 03

Deep Dives: Skills + MCP

This is the operating layer that matters most for real civil-engineering agents: Skills package reusable engineering procedure, while MCP gives the agent live access to models, files, databases, and controlled actions.

Operational Stack

Skills teach the agent how. MCP connects it to where.

For civil-engineering agents, model quality alone is not enough. The agent needs reusable playbooks for QA, naming, reporting, and review logic, and it needs live access to BIM models, inspection evidence, project systems, and structured engineering data. Skills handle the procedure layer. MCP handles the connection layer.

Progressive disclosure Reusable procedures Live system access Civil-domain grounding
Request

"Review this Revit / IFC package, flag coordination issues, and draft an engineer-ready issue log."

Skill loads

A BIM QA skill brings the naming rules, review checklist, citation format, and delivery standard the agent should follow every time.

MCP connects

MCP exposes the model, specs, issue tracker, and project database through tools, resources, and prompts instead of forcing everything into a giant prompt.

Grounded output

The result is not generic chat. It is a traceable civil workflow: cited findings, structured tickets, reusable reports, and controlled actions in real systems.

Connection Layer Protocol architecture

MCP is the live context + action surface

The Model Context Protocol is an open-source standard for connecting AI applications to external systems. In practice it turns a civil agent from a smart narrator into a system that can read models, query databases, pull files, and call approved actions in the environments where engineering work actually lives.

Tools

Model-called actions, such as creating issue tickets, running database queries, exporting reports, or launching controlled workflows.

Resources

Read-only context, such as IFC properties, design specs, sensor histories, inspection archives, or project documentation.

Prompts

Reusable templates that help users invoke structured workflows like a bridge-inspection brief or an RFI drafting pattern.

Core idea: MCP is the "USB-C for AI applications" standardization layer, so agents can connect to tools and data without custom one-off integrations for every system.
Best civil use: connect BIM, GIS, project controls, inspection databases, specifications, and sensor infrastructure without stuffing stale exports into prompts.
Key distinction: Skills teach the agent what procedure to follow; MCP gives the agent the live systems and data needed to execute that procedure.
Open-source standard Tools / Resources / Prompts Grounded live context

In civil engineering, MCP should be the way the agent reaches the model, the asset record, the code library, and the project system it must act inside.

Civil Engineering Translation

What the Skill + MCP stack looks like in the built environment

Without Skills, outputs drift away from engineering procedure. Without MCP, the agent hallucinates because it cannot see the model, the asset history, or the project system. The useful civil agent sits exactly at their intersection.

Use Case 01 BIM Coordination

BIM coordination copilot

The agent reviews federated models, finds likely conflicts, and produces structured coordination notes engineers can actually use.

Skill BIM QA checklist, naming conventions, issue severity rubric, handoff memo format.
MCP Revit/IFC access, specification library, issue tracker, project folders, and meeting notes.
Output Grounded clash or coordination report with cited evidence and optional ticket creation.
Use Case 02 Inspection

Bridge or building inspection reviewer

The agent triages inspection evidence, compares it to historical records, and drafts a defect log that follows the organization’s reporting standard.

Skill Defect terminology, severity thresholds, evidence checklist, and maintenance recommendation format.
MCP Image store, asset inventory, previous inspections, drawings, and maintenance systems.
Output Traceable defect package with ranked issues, linked evidence, and asset-specific follow-up actions.
Use Case 03 Transportation Ops

Traffic operations analyst

The agent reviews corridor conditions and supports intervention decisions instead of merely summarizing transportation data in plain language.

Skill Congestion-diagnosis playbook, KPI thresholds, narrative templates, and escalation rules.
MCP Detector feeds, signal timing databases, GIS layers, incident logs, and archive dashboards.
Output Operational diagnosis with interpretable causes, cited context, and recommended interventions.
Use Case 04 Evacuation / Safety

Evacuation and emergency planning assistant

The agent evaluates scenarios, checks facility context, and produces decision support that respects inclusive routing and safety constraints.

Skill Inclusive-wayfinding heuristics, emergency communication format, shelter and rerouting rules.
MCP Building layouts, occupancy or sensor context, weather alerts, operations plans, and contact systems.
Output Scenario-specific briefing, route plan, and action checklist grounded in the latest facility state.
Build Rules

Three design rules worth carrying into every civil-agent build

These are the high-value patterns that show up across Anthropic’s Skill guidance and the MCP documentation, and they map cleanly to engineering practice.

Rule 01

Make Skills narrow, triggerable, and explicit

Use specific task language, keep scope tight, and put critical validations near the top. A "bridge-inspection-reporting" skill is far better than a vague skill that "helps with projects."

Rule 02

Move live engineering state into MCP, not into giant prompts

Use MCP resources for model context and records, MCP tools for controlled actions, and MCP prompts for structured workflows. Do not rely on stale copy-pasted exports when the system can connect directly.

Rule 03

Let Skills orchestrate multi-step MCP workflows

For civil use cases, the best pattern is often phased execution: inspect, validate, cross-check, then write or act. Skills define the sequence and QA logic; MCP carries the data and tool calls through each phase.

Watch + Build

Useful videos and skill libraries to open right after this module

This set gives you the next layer: foundational prompting context, Claude Code workflow context, and concrete skill repositories you can study, install, or adapt.

Watch stack

Useful videos to watch

These cover prompt structure, real-world LLM workflow, Claude Code usage, and a broader framing of agentic AI systems.

Prompting Workflow habits Claude Code Agentic AI
Prompting 101 | Code w/ Claude Anthropic. Strong fit for this module because good skills start with scoped instructions and structured context. Mastering Claude Code in 30 minutes Anthropic. Useful for seeing how skills, context files, and agent workflow feel in a practical runtime. Vibe Coding - Eric Schluntz, Anthropic Team Useful as a builder-side mindset reference once you start composing skills, tools, and long-horizon agent workflows. How I use LLMs Andrej Karpathy. Useful for practical workflow habits, evaluation instincts, and how advanced users structure daily LLM work. Agentic AI: A Progression of Language Model Usage Stanford Online. Good complement here because it frames the path from prompting to tools, orchestration, and more agentic use.
Official + curated

Start with the main skill repos

If you want official templates, community curation, or places to discover reusable skill patterns fast, start with these five links.

Official repo Curated index Discovery hub
anthropics/skills Public repository for Agent Skills and the cleanest official starting point for understanding real skill structure. skillmatic-ai/awesome-agent-skills A broad curated index of agent-skill resources, useful when you want patterns beyond a single vendor stack. find-skills A focused discovery page on skills.sh for browsing portable skill packages and quickly studying how narrowly scoped skills are described. Awesome Claude Skills A community directory of Claude Code skills and plugins that is useful when you want to scan the landscape before building your own pack. AwesomeSkill.ai Another skills-discovery directory that is useful for scanning examples, comparing patterns, and finding packs to adapt into your own workflow.
Example packs

Study narrow skill examples

These examples are useful precisely because they are scoped: paper finding, research assistance, writing cleanup, and scientific workflows.

Research Writing Scientific
AI-Research-SKILLs Research- and engineering-focused skill library for turning a general agent into an AI research assistant. paper-finder A very clean example of a narrow skill: finding ML research papers instead of trying to do everything. avoid-ai-writing A focused audit-and-rewrite skill that demonstrates how sharply defined trigger scope improves reliability. Research-Paper-Writing-Skills A writing-oriented skill package for ML/CV/NLP papers, useful as a pattern for domain-specific academic workflows. scientific-agent-skills Ready-to-use scientific and engineering skills that are useful when thinking about applied research and technical analysis workflows.
Skill Guide PDF Claude Skills Blog skills.sh Ecosystem MCP Intro Tools / Resources / Prompts
Module 04

Civil AI Signal Library

Official briefings, explainers, and platform notes worth tracking if you care about agent-adjacent AI in civil engineering, BIM, digital twins, inspection, and transport operations.

Video Watchlist

Watch applied workflows, not just generic agent talk

This strip focuses on civil-engineering-adjacent workflows: Revit automation, BIM copilots, digital twins, bridge inspection, and transportation modeling. It includes several strong Rino and Caroline picks plus Autodesk, Bentley, and transportation-lab material.

8 YouTube briefings visible
Rino & Caroline BIM / MCP

How AI Is Transforming BIM: Open BIM, MCP & Revit

One of the strongest practical entries from the user-specified channel because it connects agent tooling, Revit workflows, and Open BIM rather than staying at the generic-AI level.

Revit / MCP BIM Construction
Watch on YouTube Workflow breakdown
Rino & Caroline Revit AI

Testing Revit 2027’s New AI Feature

A short, concrete check on what Autodesk is actually shipping into Revit, which is more useful for practitioners than broad future-of-AI discussion.

Revit / MCP BIM
Watch on YouTube Product reality check
Rino & Caroline Step-by-step

Revit + Claude MCP for Beginners | Step-by-Step

A practical onboarding video for model-interaction workflows that starts to look like an engineering copilot, not just a chat interface attached to BIM files.

Revit / MCP Construction
Watch on YouTube Hands-on setup
Autodesk Explainer

What Is a Digital Twin? How Intelligent Data Models Can Shape the Built World

Short but accurate context for why digital twins matter in AEC and infrastructure, especially if the page reader needs the conceptual bridge from models to operational systems.

Digital Twin BIM
Watch on YouTube Concept primer
Autodesk Construction Field Ops

How to Utilize AI in Construction | Autodesk University

A stronger construction pick than generic AI talks because it stays anchored in project delivery, data workflows, and operational uses inside the built-environment stack.

Construction Strategy
Watch on YouTube Construction workflow
Bentley Systems Bridge Inspection

Reimagining Bridge Inspection with Digital Twins

A concise bridge-inspection example from Bentley showing how digital twins and visual workflows are being positioned for real infrastructure asset management.

Inspection Digital Twin Transportation
Watch on YouTube Infrastructure inspection
ASU Transport AI Lab Research Talk

Transportation Network Modeling and Digital Twin

This adds a transportation systems perspective, showing how digital twins and open data become decision-support infrastructure rather than just building-scale visualization.

Transportation Digital Twin
Watch on YouTube Network decision support
Cadalyst Industry Briefing

Building Smarter: AI-Driven Digital Twins and the Infrastructure to Support Them

A concise industry-level summary connecting AI-driven twins to the broader data and compute infrastructure they actually require, which keeps the story grounded.

Digital Twin Construction
Watch on YouTube Short industry context
Previous
Next
Official Source Radar

Filter by workflow, not by hype

This strip is intentionally broader than pure agent demos. The useful signals in civil engineering are currently spread across AI governance, construction operations, BIM-to-twin workflows, inspection systems, and transportation asset management.

16 official resources visible
ASCE
Topic Hub

AI and Civil Engineering

ASCE’s running hub for the profession-level conversation around AI in civil engineering, including panels, articles, and member discussion about real deployment risks and opportunities.

Strategy Governance
Open ASCE hub Professional landscape
ASCE
Feature

How civil engineers can strike the AI balance

ASCE’s January 16, 2026 feature frames the practical tradeoff clearly: AI can improve traffic, inspection, and maintenance workflows, but it also raises reliability, energy, and infrastructure-capacity concerns.

Governance Strategy
Read the balance discussion ASCE Thursdays@3
ASCE
Practice

AI in civil engineering: practitioners finding their roles

Published March 25, 2026, this ASCE piece focuses on the workforce side of adoption: data hygiene, guardrails, and what human responsibility still looks like when firms start using AI systems more seriously.

Strategy Governance
Open the practice article Workforce + guardrails
Autodesk
Workflow

AI for Construction

Autodesk positions AI as workflow acceleration rather than magic: draft RFIs faster, surface risk sooner, and reduce repetitive busywork across preconstruction and field coordination.

Construction Strategy
View Autodesk workflow page AI in project delivery
Autodesk
Blueprint

Autodesk’s Blueprint for Trusted AI in Construction

A current governance reference from July 10, 2025 on data integrity, privacy, transparency cards, and why construction customers are demanding AI features they can actually trust.

Governance Construction
Read the blueprint Trusted AI guidance
Autodesk
Explainer

What is a digital twin? How intelligent data models shape the built world

Autodesk’s December 11, 2024 explainer is useful because it explicitly traces digital twins from BIM and IoT into predictive and eventually autonomous operating models.

Digital Twin Strategy
Open the digital twin explainer Concept + maturity levels
Autodesk
Class

BIM and GIS for AECO Digital Twins

An Autodesk University session focused on the BIM-to-GIS merge that underpins usable digital twins for large asset portfolios, corridors, and multi-scale infrastructure programs.

BIM / GIS Digital Twin Transportation
View the AU class AECO twin workflows
Autodesk
Lifecycle

Extending the Value of BIM to Owners with a Digital Twin

This Autodesk University article is a good owner-operator reference: it explains how BIM becomes operationally useful only when it is connected to live asset data and analytics.

Digital Twin BIM / GIS
Read the owner workflow article BIM beyond handover
Autodesk
Session

Autodesk Tandem: Delivering the Value of BIM to Owners with a Digital Twin

A practical Tandem session on digital twin maturity, outcomes, and what data structure is required before owners can actually extract value from lifecycle information.

Digital Twin BIM / GIS
Open the Tandem session Owner-operator playbook
Trimble
Panel

AI in construction: your questions answered

Trimble’s Digital Construction Summit 2024 panel is useful because it mixes optimism with industry concerns, including legal risk, implementation reality, and governance across the built environment.

Strategy Construction Governance
Open the Trimble panel Digital Construction Summit 2024
Trimble
Outlook

AI for Construction

Trimble’s January 3, 2024 outlook is operationally grounded: centralize project data, ask better questions, and do not trust AI outputs blindly in a safety-critical industry.

Construction Strategy
Read Trimble’s outlook Practical setup advice
Trimble
Guide

Why BIM Needs Digital Twins

A straightforward Trimble guide explaining the handoff from BIM to connected operations, with IoT as the link that turns model data into a living digital twin.

Digital Twin BIM / GIS
Open the Trimble guide BIM meets IoT
Bentley
Platform

Infrastructure Digital Twins

Bentley’s overview is still one of the clearest infrastructure-side summaries of the twin model: combine engineering, reality, geospatial, and IoT data to improve decisions across design, construction, and operations.

Digital Twin Transportation Strategy
Explore Bentley twins Infrastructure lifecycle view
Bentley
Bridge Case

America Has 600,000 Bridges. Engineers Using AI Just Found a Better Way to Inspect Them

Published March 25, 2026, this Bentley case story shows drones, digital models, and AI analysis being used to inspect a historic bridge faster, more safely, and with less field exposure.

Inspection Transportation
Read the bridge case AI-guided inspection workflow
Bentley
Forum

Bentley brings AI and digital twins to the TRB annual meeting

Bentley’s January 12, 2026 TRB briefing ties AI and connected data directly to transportation policy, research, and investment rather than treating digital twins as isolated software demos.

Transportation Strategy Digital Twin
Open the TRB briefing Transport policy + twins
Bentley
System

AssetWise Inspections

A field-to-office inspection system for bridges, tunnels, roads, rail, and transit assets, with a concrete view of how digital inspection workflows become operational systems instead of slideware.

Inspection Transportation
Explore AssetWise Inspection operations stack
Previous
Next
Module 05

Podcast Signal Room

Filter Episodes
6 Listen Notes episodes visible
Fire Science Show · Your Talk

Setting up your own chatbot

Agentic AI RAG Local LLMs
Open Episode
Fire Science Show · Mar 2026

241 - Opportunities with AI (in 2026) with MZ Naser

Agentic AI Civil Engineering Guardrails
Open Episode
The ConTech Crew · Apr 2025

Cutting construction design time by 80% with AI

Construction Design Automation Workflow Change
Open Episode
Future Construct · Oct 2025

BIM and AI in Construction

BIM Construction Digital Twin Adjacent
Open Episode
The Site Visit · Mar 2024

How AI and Diversity Are Revolutionizing the Construction Industry

Construction Applied AI Business Ops
Open Episode
InfraTalk America

Building the Future: Digital Transformation in Infrastructure and Transportation

Infrastructure Transportation Digital Transformation
Open Episode
Previous
More episodes
Search Engine Source

More Audio & Learning Resources

Rino & Caroline

BIM, Revit, and MCP workflows

AI for CE

Civil engineering AI channel

DeepLearning.AI

Andrew Ng's AI agent courses

HuggingFace Papers

Daily AI research papers

AI in Transportation Lab

Texas State University

arXiv cs.AI

Latest AI research
Module 06

Open-Source Tools & Repos

A broader repo shelf: civil benchmarks, BIM workflows, agent systems, MCP infrastructure, and applied safety research worth opening before the framework docs.

Eason-Li-AIS/DrafterBench

Benchmark suite for civil engineering drawing automation, useful when you want to test whether an agent workflow actually improves drafting tasks instead of just demoing well.

Benchmark Civil AI Evaluation
mac999/BIM_LLM_code_agent

A BIM-focused code agent prototype that ties LLM prompting to model-side work instead of leaving the assistant detached from project geometry and metadata.

BIM AEC Code Agent
Danitilahun/n8n-workflow-templates

Reusable n8n patterns for approvals, extraction, notifications, and orchestration when you want agent behavior without hand-coding the whole automation stack.

Automation Workflow n8n
poloclub/transformer-explainer

Interactive transformer visualization for teaching attention, token flow, and model internals before students start treating agents like a black box.

Education Transformers Interpretability
NousResearch/hermes-agent

Extensible Nous Research agent stack positioned as “the agent that grows with you,” built around tools, skills, plugins, and multi-surface workflows.

Agent Stack Skills Plugins
lsdefine/GenericAgent

Self-evolving agent repo that grows a skill tree from a compact seed and targets broad system control with lower token usage.

Self-Evolving Skill Tree Systems
bytedance/deer-flow

ByteDance’s long-horizon SuperAgent harness for research, coding, and creation with memories, tools, skills, sandboxes, and subagents.

Long-Horizon SuperAgent Sandbox
microsoft/agent-lightning

Microsoft’s open-source agent training repo, useful when you want to improve agent behavior beyond prompt tweaking and static workflows.

Training Agents Microsoft
googleapis/mcp-toolbox

Open-source MCP server for databases from Google APIs, useful when agents need structured access to SQL backends through MCP.

MCP Databases Infrastructure
pozapas/Crowd-Analyzer

Computer-vision crowd mobility analyzer with YOLO, Kalman tracking, PedPy plots, and pedestrian-vehicle interaction analytics.

Computer Vision Mobility Safety
pozapas/traffic-safety-vlm

Research-grade multimodal traffic safety app for live streams and recorded video, built around vision-language model analysis.

VLM Traffic Safety Video
pozapas/neuroevac

Research-grade Streamlit dashboard for EEG analysis in evacuation scenarios, with preprocessing, mapping, anomaly detection, clustering, and AI reporting.

EEG Evacuation Streamlit

Key Agent Frameworks & Libraries

Core orchestration libraries, kept separate from the domain-specific repos above.

LangChain

Agent orchestration

CrewAI

Multi-agent teams

AutoGen

Conversational agents

LlamaIndex

RAG framework

DSPy

Declarative LM programming

Transformers

Model backbone
About Me

Amir Rafe

Postdoctoral Researcher at Texas State University, AI in Transportation Lab.
Research in causal AI, computational epistemology, complexity, and safety.

My work focuses on how intelligent systems build and apply knowledge in dynamic, safety-critical environments — with emphasis on pedestrian dynamics, evacuation science, and causal reasoning. I integrate AI, human behavior modeling, causal inference, and simulation to design decision-support tools that are robust, explainable, and inclusive.