Fifteen million people use the agents I ship.
I started Simplified in 2020. Six languages, thousands of paying customers, revenue tripling year over year, and a team of more than 50. Other agents can use it too, through the Simplified MCP server.
├ now2020 – todaySimplified: founder and CEOok
├ nowthe systemRiley, a marketing agent in productionok
One orchestrator, the specialists it hands work to, and a person it knows when to ask.
On every model callWorkspace context and the brand kit are injected.
OrchestratorPlans with a todo list it writes itself. One pinned model, 33 skill folders.
task, and waitstart, check, cancel
Sub-agent, synchronousCreative producer. One item per call; the parent waits.
Sub-agent, asynchronousBrand book builder. Its own run; the parent polls and can cancel.
Tools35 integration providers behind an MCP gateway.
MemoryWorkspace-scoped, consolidated after each turn.
StateA Postgres checkpoint per run: pause, resume.
PeopleTen ways to stop and ask.
├ nowthe proofhow it is evaluatedok
Evaluation cases it has to pass before it ships.
Each one runs the real agent over its real interface, including the pause for a person’s approval. About twenty kinds of exact check, plus rubric scoring by a judge model.
Open source
Probe
I also built a framework that evaluates agents the way a user meets them: in a real browser, with a screenshot and a reason for every check.
github.com/celeryhq/probe├ nowthe lessonswhat production taught usok
Five things running agents in production taught us.
Models fill optional arguments with placeholders.
An ads API rejected a call that set two budgets at once. We now strip empty optionals before any tool call.
Twenty records are a job for code, not prose.
A run created campaign items with fields missing. A typed tool now expands and validates them, and the model is not allowed to create them directly.
A missing tool has to fail the build.
When a toolkit silently failed to register, the model invented substitutes. Now it is an error before anything runs.
Prove the pipeline in plain code first.
Only then turn it into a skill the agent follows.
No inspection record, no delivery.
An image cannot ship unless a check was actually logged for it.
├ nowthe stackwhat is open every dayok
I work AI-native. This is what I have open every day.
- build
- Claude Code, Codex, Devin, Replit
- design
- Claude Design, Paper
- think and research
- Claude, Perplexity
- agents in production
- LangGraph, deepagents, MCP
- evaluate and trace
- Probe, Langfuse, Playwright
- images and video
- Simplified, HeyGen, HyperFrames
├ recall2020 back to 2005what came beforeok
Before it had a name, I was already doing the parts.
├ recall2020emailed the CTO of OpenAIok
I emailed the CTO of OpenAI. ChatGPT was still two years away.
I had been building with language models since GPT-2, so I wrote to the person running the technology.
From: KD Deshpande
To: Chief Technology Officer, OpenAI
Sent: 2020
The full text of the email will go here.
├ recall2020Meta: lead product manager, Messengerok
Product decisions at 1.3 billion people a month.
I led product strategy for Messenger, and learned how engineering, design and data science agree on anything at that size.
├ recall2017 – 2020Uber: senior product manager, growth and monetizationok
Ranking with real money on the line.
Uber had no advertising business. I wrote the case, started it, and launched Sponsored Listings on Uber Eats: experimentation, personalization and monetization at the moment someone decides what to order.
├ recall2014 – 2017Adobe Marketo: head of mobile product, then general managerok
Owning the number, not just the roadmap.
At Adobe Marketo I ran mobile product, then became general manager of innovation and partnerships, reporting to the chief product officer. Push, in-app, SMS, location, and a launch in Japan.
├ recall2011 – 2014Vessel.io: CEO and co-founderok
Measuring which version is better, in 2011.
I co-founded Vessel.io, an A/B testing and analytics platform for mobile apps, wrote its SDKs, took it through 500 Startups and sold it to Marketo, now Adobe Marketo. It is the same question I ask of agents now.
├ recall2011 – 2012TuneIn: senior software engineer, mobileok
Shipping to hardware I did not control.
I built mobile radio for millions of daily listeners, on phones and in dashboards, alongside Ford, Tesla, HTC and Samsung.
├ recall2005 – 2010engineer: Symantec, BMC Software, co-founder and CTO at Virtuonok
Where I learned how systems break.
Two computer science degrees, then Symantec and BMC, then my first company, Virtuon, where I ran the technology.
└ nextopenwaiting on your message…
That is the run so far. If you are building with models and agents, write to me.
Koustubha Deshpande, San Francisco Bay Area. This page was built by directing an agent in Claude Code, over eight versions. The prints were generated through the Simplified MCP server.