Work/AI Agent Engineering
01 / 05
Case Study · AI Agent Engineering

Designing the canvas after it grew a mind.

For years I made the canvas easier to draw on. Then agents arrived — and the canvas wasn’t holding flows you draw anymore, it was holding things that act on their own. This was the net-new product where I redesigned the canvas around that shift, authored the framework behind it, and made one call against the whole field: I turned the flow top to bottom. Shown at the Informatica World 2025 CEO and CPO keynotes.

Role
Lead UX — 0→1
Surface
Net-new agentic builder
Partners
Product & engineering, in their lanes
Shown at
IW 2025 CEO + CPO keynotes
00Hiring manager debriefDebrief 01Context 02Role 03Approach 04Solution 05Results 06Impact
TL;DR · the 60-second version

What happened here, before you scroll.

Agents changed what a canvas is for. I started from a question nobody had a clean answer to — when the canvas holds agents that act, not flows you draw, what should it look like? I authored CHESS to answer it, and made the keystone bet: flip the orientation top-to-bottom, backed by research. A net-new product — 347 → 1,938 preview registrations in 30 days, 630+ weekly active users. Here’s the whole thing — jump in anywhere.

The opportunity
A net-new agentic builder: let people build, connect, and govern intelligent workflows without thinking in code.
see context →
My role
Led UX end to end — research, the CHESS framework, the interaction model, the IW’25 narrative. Product & engineering in their lanes.
see role →
How I got there
Authored CHESS, ran the orientation trade-off against five tools and Process Diagrammer’s own research, then designed the structure-and-delegation model.
see approach →
What I designed
A top-to-bottom agentic canvas — planning agent, tools, inline agents, supplemental text, co-pilot — built around CHESS:
see solution →
Where it landed
347 → 1,938 registrations in 30 days, 630+ WAU. Launched at the IW’25 CEO + CPO keynotes; Yahoo Finance & BusinessWire coverage.
see results →
Why it matters
This is where the canvas I’d specialized in grew a mind — and set up everything that came after.
see impact →
01 — Context

The canvas grew a mind.

Process Diagrammer made the canvas usable at scale — the surface where people draw enterprise workflows, from loan approvals to real-time fraud detection. Then agentic AI arrived. Agents could reason, call tools, and act on their own, not just sit as nodes you wire by hand. Being agentic means more than wrapping APIs in an MCP — it means applying agentic principles across an asset’s whole lifecycle.

The problem
When the canvas holds agents that act on their own — calling tools, making decisions — a left-to-right flow built for drawing breaks down. People need to see the orchestration and stay in command of it.

What changed is capability: agents got good enough to plan, reason, and call tools on their own — so the canvas’s job moved from drawing a flow to orchestrating agents. This was the net-new product built for that shift.

Before
The canvas you draw
Nodes wired by hand, a flow that just runs. Powerful, but it only ever executes what you laid out.
Now · net-new
The canvas that thinks
Agents reason, call tools, and act — the person orchestrates and stays in control, without thinking in code.

The stakes and scale are real: these workflows run regulated, high-volume processes — from loan approvals to fraud detection, scaling to 10,000 runs a minute. Designing a canvas where a human stays in command of agents that act on their own is the problem I set out to solve.

↑ back to the debrief
02 — Role

What was mine, and who I worked with.

I led UX end to end on the net-new agentic builder — the research, the CHESS framework I authored, the interaction model, and the narrative shown at the IW’25 CEO and CPO keynotes. Product and engineering owned scope, feasibility, and delivery. I name the split plainly because an honest account is more useful than a flattering one.

What I owned

The CHESS framework · the top-to-bottom orientation decision · the structure-and-delegation interaction model · the IW’25 keynote narrative · prior ownership of the Process Diagrammer canvas this built on.

What I worked against

No precedent — no one had designed an agentic canvas · a convention (left-to-right) I chose to break · a net-new 0→1 carrying a hard keynote date.

↑ back to the debrief
03 — Approach

I built the framework first, then made the bet.

Three moves: a framework to reason by, the one orientation call that defined the canvas, and a structure for how humans and agents share the work.

1 · CHESS — the five principles I authored

I needed a way to reason about agentic interaction before drawing a single node. So I wrote one. CHESS gave the team a shared language for every decision that followed.

2 · The orientation bet — top to bottom, against the field

Every node canvas I knew ran left to right. I turned this one top to bottom — the riskiest call in the project, and the one I’m most sure of. The reasons stacked up: horizontal scrolling is disliked; large processes stay linear downward instead of sprawling sideways; left-to-right hits a width wall and wraps, breaking focus; and top-to-bottom mimics how real work reads — code, docs, and configs all run down the page.

ToolOrientationRead
ServiceNowLeft-to-right (BPMN)Standardized, but stagnated — lacks structure
Flowise AIMulti-directionalAnywhere & everywhere — messy at scale
n8nLeft-to-rightAutomation engineers
OnDemand.ioLeft-to-rightRAG / LLM teams
BuildShipTop-to-bottom ★Enterprise product teams — the match
The research backed the bet before I made it: in Process Diagrammer’s own study, only 10% of users wanted to keep left-to-right over any other change. BuildShip — top-to-bottom, built for enterprise teams — was the one tool that fit the work, and it confirmed the direction.
3 · Structure & delegation — how humans and agents share the work

A planning agent is the orchestration brain: it runs at least once and, from the prompt, decides which tool or inline agent to invoke. Tools and inline agents sit on the right — pure delegation for execution. Supplemental text sits on the left — prompts, parameters, and course corrections that enrich the agent’s reasoning. A co-pilot surfaces suggestions and error-handling right where the user needs them. Two entry points keep the user in command: drop a tool on the main flow for a deterministic A→B→C, or delegate it to the planning agent and let the prompt decide the sequence.

What surprised me: breaking the left-to-right convention felt reckless — until the data backed it. Only 10% wanted to keep it. The lesson stuck: a strong convention isn’t the same as a strong preference.

↑ back to the debrief
04 — Solution

A canvas built to orchestrate, not just draw.

The model lets a person build, connect, and govern agentic workflows without thinking in code — and stay in command of agents that act on their own. Five parts, at a glance:

01 · Orientation
Turn the flow top to bottom.
The sequential flow runs down the page, not across it — no horizontal sprawl, no width wall, no wrapping. Large agentic processes stay linear and readable.
For the user
A long workflow reads like a document, in one predictable scroll.
Why
Horizontal canvases break focus at scale; downward mimics how real work reads.
Frame coming
02 · Structure
A planning agent as the brain.
A planning agent runs at least once and, from the prompt, decides which tool or inline agent to invoke — orchestration sits in one clear place instead of scattered across nodes.
For the user
You read the plan first, then the steps — intent stays legible.
Why
Orchestration is the new hard part; it deserves a home, not a guess.
Frame coming
03 · Delegation
Tools and inline agents, on the right.
The right side is pure delegation for execution — tools and inline agents that run only when the plan calls them. Two entry points: drop a tool on the main flow for a deterministic A→B→C, or delegate it to the planning agent and let the prompt decide.
For the user
Choose certainty or flexibility per step, without micromanaging.
Why
Not every step should be fixed, and not every step should be loose.
Frame coming
04 · Context
Prompts and corrections, on the left.
Supplemental text lives on the left — prompts, parameters, and course corrections that enrich the agent’s reasoning — expanding only when needed. Metadata and a co-pilot keep help where the user is looking.
For the user
Add context where it belongs, without cluttering the flow.
Why
An agent is only as grounded as the context it can see.
Frame coming
05 · Oversight
Keep a human in command.
No-code UX, co-pilot checkpoints, and collapse / expand across execution levels keep a person oriented as workflows grow — in control of agents that act on their own, not buried by them.
For the user
Collapse the whole thing, or open just the level you need.
Why
Human-in-the-loop is the ‘H’ in CHESS — control is a feature, not an afterthought.
Frame coming
One canvas, two ways to drive — deterministic when it must be, delegated when it can be.
↑ back to the debrief
05 — Results

What shipped — with numbers.

This one has numbers, and they’re verifiable — a net-new product, launched at a global keynote, that grew quickly in preview.

Adoption
347 → 1,938 preview registrations in 30 days — a 5.6× jump — and 630+ weekly active users.
Stage
Launched at the IW’25 CEO + CPO keynotes, with coverage in Yahoo Finance and BusinessWire.
Influence
CHESS became the shared language for agentic interaction on the platform, and the top-to-bottom canvas set the direction for what came next.
↑ back to the debrief
06 — Impact

What it changed, and what I’d carry forward.

This is where the canvas I’d specialized in grew a mind. Design’s job moved from drawing flows to orchestrating trust between a person and the agents acting on their behalf — and it set up the headless work that followed, when the canvas disappeared entirely.

What I’d do differently: V1 capped at three levels of execution to make the keynote date — I’d push for deeper nesting sooner. And I’d put the orientation flip in front of more enterprise hands before the launch, not just after.

The canvas didn’t get smarter. The work did — and design’s job became keeping a human in command of it.
Note: this case study describes concepts, research observations, and design decisions only — no customer names, no demo data, no unreleased product specifics. Competitor tool comparisons reflect publicly documented behavior.
↑ back to the debrief