Talk · Presented at Google · July 21, 2026

What is the best framework for FDE work?

“Framework” has two meanings. This talk covers both: the decision framework you need on-site to win the deal, and the code framework that lets you deliver it in three weeks. Field notes from two weeks in Hangzhou and two signed enterprise customers.

Get started in the docs Jump to the live demo Slides (.pptx)

Part 1 · Sales

The deal is decided on-site

Every company has people doing repetitive work on a screen. Those workflows are too long-tail and too custom for any SaaS to ever cover — which makes the market for screen automation effectively unlimited. But turning that into signed deals demands three rapid judgments, made in the room:

Is this company worth it?

Is the customer a young, growing company? That decides whether the deal compounds or dead-ends.

Requires: Judgment of people and companies

Can we finish in three weeks?

Can the requirement be shaped, on the spot, into something deliverable in three weeks?

Requires: Product judgment

Can today’s AI actually do it?

Is the ask within what current models can reliably deliver — not in a demo, in production?

Requires: Technical judgment

These three judgments rarely live in one employee. Early on, this isn’t a job for an FDE — the founder has to go in person.
Every early deal is a product-direction decision. FDE teams exist to replicate a validated playbook — not to blaze the first trail.

Part 2 · Tech

Promising three weeks takes a cadence

Week 1

Trade a demo for trust

Converge the requirement on-site and ship a running minimal demo the same week.

Week 2

Wire up real systems

Connect the customer’s internal systems and real data. Handle the edge cases.

Week 3

Hand over and sign off

The customer’s engineers can read it, change it, and maintain it. Only then is it delivered.

The cadence holds because you are not building an app. ConnectOnion is a runtime platform for skills: the frontend, backend, terminal, browser, and file system are already there. What you write is the skill — prompts and scripts:

$ pip install connectonion

agent = Agent("assistant", tools=[my_tool])
agent.input("automate this workflow")

$ co email

Email

Every agent gets its own email address, automatically. Gmail and Outlook ship built in — send, read, search, even schedule mail from the terminal.

$ co browser

Browser

Screenshots, scraping, automation — one command. In code, a real stateful browser: start, navigate, click, fill.

$ co ai

Terminal

bash and Shell are built-in tools with an approval system, and co ai puts a framework-native coding agent in your shell.

Three skills ship in the box — email, browser, terminal — covering most of what a screen-automation deal needs. And when it’s done, co deploy ships it to the cloud.

Model-agnostic

OpenAI, Claude, Gemini, Ollama — the customer picks the model, not you. Switching is one string.

Readable code

No thick abstraction layers. The customer’s engineers understand the code the first time they open it.

MIT open source

Everything is public on GitHub. Security review isn’t a negotiation — they just audit it.

Python / TypeScript / Rust

Deliver in whatever stack the customer already runs.

A runtime for skills

You don’t build an app around the agent. You write a skill — prompts and scripts — and deploy it onto the platform.

Everything else is already there

Frontend, backend, command line, browser, file system — shipped with the platform, not assembled per project.

A tiny, readable core

The core is genuinely small and flow control is plain. About five concepts cover the whole platform, and permission control is built in.

Test in Claude Code or Codex, then ship

Build and test your skill with a coding agent, then deploy it straight onto the platform — that’s the whole FDE loop.

Live demo · Zero to running

Usable now. Deployable now.

You don’t assemble a stack. Install the platform, write the skill, deploy — the same skill you demo is the one that ships.

1 · Install

$ pip install connectonion

One package brings the whole platform — frontend, backend, terminal, browser, file system.

2 · Write

agent = Agent("assistant",
              tools=[my_tool])
agent.input("automate this")

A skill is prompts plus plain Python scripts — no wrappers, no registries, no app to build.

3 · Run & deploy

$ python agent.py
$ co deploy

It runs anywhere Python runs — and one command deploys it to ConnectOnion Cloud. Nothing changes in between.

Follow along in the docs Source on GitHub

Conclusion

The best FDE framework = a decision framework for the room + a code framework that disappears into the handover.

Three judgments decide whether to take the deal. Two lines of code make sure you can deliver it. The first still takes a founder in the room — the second is already open source.