← Talks
Agents in Practice
From overlays to agentic AI
Zihao Zhang, Ph.D.
Associate Professor of Landscape Architecture
City College of New York

Who’s in the room?

Four quick questions, to know the room.

  • Hands up if you use GIS in your work or studies.
  • Hands up if you use Grasshopper or any scripting.
  • Hands up if you have used a chat assistant for work or school.
  • Hands up if you have used an agent that edits files for you.

By the end of this session, you will be able to

  • Trace how computational thinking has appeared in design practice, from overlay mapping to GIS, CAD and parametric modeling, and consider where AI agents sit among these tools.
  • Explain how an agent differs from a chat assistant, and how MCP connects an agent to software you already use.
  • Apply folder structure and a written context file to control what an agent knows across working sessions.
  • Evaluate which tasks in your own work can be delegated to an agent, weighing verification, confidentiality, cost and time.
Part 1

Computational thinking in design, a short history

Part 2

How agents work, and MCP

Part 3

Context through folder structure

Part 4

Discussion

Part 5

Try it yourself

Charles Eliot — Sudbury Valley survey map
Warren Manning — Billerica Town Forest map
Jaqueline Tyrwhitt — sieve map overlay, 1950
Ian McHarg — Design with Nature layer-cake overlay diagram
SYMAP at a mainframe console
SYMAP line-printer map, 1963
Ivan Sutherland drawing on the Sketchpad screen with a light pen at the TX-2 console, 1963
Six Sketchpad stills: a light pen draws lines that the program straightens, constrains and copies
An Edson bronze spline weight (duck) resting on lofting drawings
A row of ducks fairing a wooden spline batten across hull lines
B-SPLINE CURVE P0 P1 P2 P3 P4 control polygon (dashed) + smooth curve NURBS SURFACE
FROM B-SPLINE TO GRASSHOPPER 1967 “B-spline” named Schoenberg coins the term 1972 Cox–de Boor recurrence makes B-splines computable 1973 Riesenfeld thesis → Utah with Elaine Cohen · geometry lab 1975 Versprille thesis non-uniform + rational form 1980 NURBS named Boeing TIGER · R. Blomgren 1984 AGLib kernel Applied Geometry (Blomgren) 1994 McNeel licenses AGLib Sculptura plug-in · M. Gibson 1998 Rhino 1.0 Robert McNeel & Assoc. 2007 Grasshopper David Rutten · visual programming
IBM PC, 1982
Man using AutoCAD on a PC, 1987
Architecture office before the 1980s
Computers at NASA / JPL-Caltech
Martha Schwartz, Jacob Javits Plaza, New York, 1997
James Corner Field Operations, Philadelphia Navy Yards Central Green, 2015
James Corner and Alex MacLean, Taking Measures Across the American Landscape, 1996
James Corner and Alex MacLean, Taking Measures Across the American Landscape, 1996
Bradley Cantrell, Digital Drawing for Landscape Architecture, 2010
Diana Balmori, Drawing and Reinventing Landscape, 2014
Nadia Amoroso, Representing Landscapes: Digital, 2015
Zaha Hadid
Zaha Hadid
Z+T Studio
Z+T Studio
Z+T Studio, parametric construction
Z+T Studio, parametric construction
Philip Yuan, Chi She, robotic construction
Philip Yuan, Chi She, robotic construction
Philip Yuan, Chi She, robotic construction
Philip Yuan, Chi She, robotic construction
Fletcher Studio, South Park
Fletcher Studio, South Park parametric analysis
Fletcher Studio, South Park parametric analysis
Plasma Studio and Groundlab, Xi'an Expo Flow Garden
Xi'an Expo Flow Garden detail
Xi'an Expo Flow Garden detail
Little Island, New York, 2021
Little Island, New York, 2021
Little Island, New York, 2021
Little Island, New York, 2021
Field Operations and Diller Scofidio + Renfro, The High Line, 2009
Bradley Cantrell — sensing and responsive landscapes
Timeline of landscape representation tools from Sketchpad to web maps, grouped as drafting, analysis, modeling and communication, fading into “Agentic AI?”

“Scores communicate but do not control.”

Lawrence Halprin’s hand-drawn RSVP Cycles diagram
Part 1

Computational thinking in design, a short history

Part 2

How agents work, and MCP

Part 3

Context through folder structure

Part 4

Discussion

Part 5

Try it yourself

A subway-map path beside the line “One prompt, job done.”
“Slides in 10 minutes. Not 10 PM.”
GPT-generated presentation board, waterfront wetland
GPT-generated presentation board, seasonal wetland
GPT-generated presentation board, post-industrial landscape
GPT-generated presentation board, access map

A chatbot is a person in a locked room

Chat assistantAgent
InteractionOne prompt, one responseBreaks a task apart, acts, checks, repeats
ReachText out; no hands, no eyesFiles, a terminal, a browser, connected software
MemoryWhat you pasted inThe folder it works in
Your roleWord each question wellDelegate a goal; review the result
MCP is the key to the door.
THE AGENT LOOPYour taska goal, not a sentenceReadfiles, context, the goalPlanbreak it into stepsActedit, run, call a toolCheckdid that work?Ask youunsure, or needs approvalDonereports what changedLoops untilthe check passesor it needs you
WHAT AN AGENT CAN TOUCHAgenton your computerFilesthe project folder, standardsTerminalbatch-rename 400 site photosBrowserlook up a code sectionConnected softwareRhino, QGIS, a spreadsheetYou decide which doors open,and whether it asks first.
MCP — ONE STANDARD PLUGAgentany MCP clientMCPadapterRhinoyour 3D modeladapterQGISsite dataadapterBlenderrendersadapterSpreadsheetplant scheduleAn open standard,published late 2024.The adapter runs onyour machine; what theagent reads goes tothe AI provider.Some adapters areexcellent; many areprototypes.
The request typed to the agent and the start of its work log: build a landform, run a slope analysis, capture an isometric view
Isometric view of the landform, colored green, yellow and red by slope band

Several companies make coding agents

  • Claude Code (Anthropic), Codex (OpenAI), Cursor (Anysphere), Gemini CLI (Google), and others. All of them work inside a folder on your computer.
  • They share open standards: MCP to connect software, AGENTS.md for a project’s instructions.
  • The examples today use one of them. Naming them is not an endorsement.

Check-in: what would you want an agent to open for you first?

Part 1

Computational thinking in design, a short history

Part 2

How agents work, and MCP

Part 3

Context through folder structure

Part 4

Discussion

Part 5

Try it yourself

WHAT THE AGENT KNOWSNot in the windowlast week’s sessionthe shared drivewhat you meantthe drawing setyour phone callCONTEXT WINDOWIts instructionsAGENTS.md, loaded every sessionFiles it has openedwhat it read in this sessionThe conversationsummarized when it runs longClose the session:the conversationis gone.The files stay.Anything outside the window is not there for the agent.

Two kinds of memory

The conversationThe folder
LastsThis sessionAs long as the project
Use it forSteering today’s workDecisions that should outlive today
Everyday equivalentA phone callThe project folder, the standards binder
Said it twice? Ask the agent to write it into the file.

The project’s standards manual

  • The agent reads AGENTS.md at the start of every session (some tools read CLAUDE.md).
  • Tell the agent what to add, and it writes the file.
  • Keep it short, and ask the agent to audit it now and then.
AGENTS.md
# Riverside Studio website

Small LA practice, 8 people, Queens.
Plain HTML, published on GitHub Pages.

## Folders
_source/    photos + text from the team;
            read, never publish
projects/   one page per project
assets/     images, built from _source/

## Rules
Images: long edge 1920 px.
Office facts live in office.md.
Ask before deleting anything.
✓Clear names: read at a glance
website/
├─ AGENTS.md
├─ office.md
├─ index.html
├─ projects/
│   ├─ riverside-park.html
│   └─ astoria-schoolyard.html
└─ _source/
    └─ riverside-park/
        ├─ 01-plaza-dusk.jpg
        └─ description.md
✗Vague names: opened one by one
website/
├─ notes.txt
├─ page3_final_v2.html
├─ page3_final_v2 (1).html
├─ NEW/
│   ├─ IMG_4471.jpg
│   ├─ IMG_4472.jpg
│   └─ Untitled-1.html
└─ old stuff/
    └─ use this one.docx
Single source of truthCOPIES DRIFTPOINTERS DO NOTindex.html212-555-0142contact.html212-555-0142footer212-555-0198 (old)Three copies, one already wrong.office.mdphone:212-555-0142index.htmlsee office.mdcontact.htmlsee office.mdfootersee office.mdproposal templatesee office.mdOne edit fixes every page.

Designer websites, built with Claude

  • zihaozhang.cc: projects, CV, teaching, talks. One person, no web designer.
  • bradleycantrell.com: Bradley Cantrell, landscape architect and scholar.
  • Next: how a site like this is set up, once, and the daily routine.
The projects gallery on zihaozhang.ccThe home page of bradleycantrell.com
Gitrevisions, on your computerGitHubshared plan room, onlineGitHub Pagesyour public websiteYOUAGENT01Install Git+ your agentone-time download02GitHubaccountturn on two-factor03Organizationfor a shared site04Sign inthis computernormal terminal05Createthe repothe online folder06Uploadthe foldera dated version07Turn onPageslive in a minute08Own domainoptionalStuck at any step? Paste the error to the agent and ask it to walk you through.
EVERY DAYYou ask“Add the Riversideproject.”Agent editsand previews, followingAGENTS.mdYou reviewapprove it,or correct itAgent publishesthe site updatesin a minutenext changeCorrect the same thing twice? Ask the agent to write the rule into AGENTS.md.The daily loop is where the context file grows.
Stacked subtitle meme of Rick Rubin on 60 Minutes: “Do you play instruments?” “Barely.” “Do you know how to work a soundboard?” “No.” “So what are you being paid for?” “The confidence that I have in my taste and my ability to express what I feel.”
THREE KINDS OF KNOWLEDGEButtonsWhich command?Patch, Offset, Remapwhich menu, which inputspecific to one programAN AGENT CAN DO ITBehaviorWhat will it do?where water drainshow a slope readshow a line weight printsYOU, WITH AN AGENT CHECKINGJudgmentIs it right, here?for this sitefor these peoplein this climateYOUTo check what an agent made, you need to know how the thing behaves.
Part 1

Computational thinking in design, a short history

Part 2

How agents work, and MCP

Part 3

Context through folder structure

Part 4

Discussion

Part 5

Try it yourself

TOOLS PER PHASE — WHERE DOES THE TIME GO?1Site Research& AnalysisGISPhotoshopIllustrator2ConceptualDesignRhinoCADRenderingPhotoshopTrace paper3DesignDevelopmentCAD4ConstructionDocumentationCAD???proposalsplant schedulesCAD cleanupportfolio pagesthe websiteEvery hand-off between these boxes costs someone an afternoon.
Part 1

Computational thinking in design, a short history

Part 2

How agents work, and MCP

Part 3

Context through folder structure

Part 4

Discussion

Part 5

Try it yourself

Try it yourself

Start now if you have a laptop, or later at your own desk. Each guide walks you through, step by step.

QR codeRhino + MCPSlope bands, a layer audit, a plant scheduleOpen the Rhino guide
QR codeFolder contextGitHub setup, then a context file the agent writes and auditsOpen the folder-context guide

Thank you.

/ 45

Agents in Practice

Talk · 2026

A talk for landscape architects, designers and students on AI agents: tools that work inside a project folder, read and edit its files, and connect to software you already use. It starts from a claim about the profession. Landscape architects have always done computation, first on layers of mylar and later in GIS, CAD and Grasshopper, and almost none of those tools were built for the discipline. Agents are the first tools a designer can reshape around their own work by describing what they want.

The talk looks at computational thinking in the design professions, from Eliot’s and McHarg’s overlays through SYMAP, AutoCAD and parametric modeling; explains how an agent differs from a chat assistant and how the Model Context Protocol connects it to programs like Rhino; and shows how folder structure and a short context file, written and maintained by the agent itself, carry a project across working sessions. It closes Part 3 with the workflow behind this website, from setup to the daily routine. A discussion follows on what designers would hand over, what they would keep, and the conditions an agent must meet before it touches client or confidential files. The last part points to two hands-on guides, for anyone who wants to try it in the room or later.

Two step-by-step guides continue the hands-on work:

An agent inside RhinoRhino + MCP: slope bands, a layer audit, a plant schedule Folder context and a website on GitHubGitHub setup, then a context file the agent writes and audits