From Skill to Agent: When a Text File Isn’t Enough
April 21, 2026
Adam Lewis Senior Software Engineer A coworker of mine built a Go CLI for the Harvest time-tracking API. It’s a solid tool, but every time I wanted to use it
They had ambition, and it showed.
This team of engineers had built something remarkable: intelligent agents that could parse documents in seconds, respond around the clock, and surface insights from oceans of unstructured data.
But there was a catch.
It didn’t scale.
The agents worked beautifully in demos. The models performed under test conditions. But when it came time to embed them in secure enterprise environments, the system began to crack.
Security teams had no visibility into what the agents were doing behind the scenes.
The team had built an engine with enormous potential—fast, powerful, endlessly capable. But the frame couldn’t support the speed.
That’s when they called us.
You could sum it up in one work: fragility.
We engineered a headless Jupyter extension—no UI, no friction—so agents could run invisibly and securely. That single change enabled them to operate safely inside the company’s AI platform, without interrupting core workflows.
By stripping unnecessary loading behaviors, we cut cold-start times dramatically. That seven-second lag? Gone.
We brought R support to full parity with Python using RPy2 and IRKernel. Analysts no longer had to switch languages or hack together workarounds.
We re-architected the SDK using Fern, hardening its interfaces and standardizing data handling across Arrow, Parquet, Matplotlib, and Plotly.
We customized and compressed the Jupyter UI to sit inside the company’s multi-tabbed interface. The notebooks became the platform.
We didn’t ask this team to change how they operate. We met them where they were—and gave their agents the infrastructure to match their intelligence.
We made their product faster, safer, more auditable, and easier to control because we listened first.
| Capability | Before | After |
|---|---|---|
|
Execution
|
Tied to Ul, fragile |
Headless agent runtime inside Jupyter |
|
Startup Time |
7-second delay |
Cold-start optimized |
|
R language Support |
Incomplete, unstable |
Full parity with Python |
|
SDK + APIs |
Manual, error-prone |
Streamlined, scalable with Fern |
|
Visualization |
Visualization |
Isolated charts |
|
UI/UX |
Stock Jupyter |
Embedded, branded Athena interface |
Core agent workspace
Secure, containerized execution
First-class R support
Robust SDK tooling
Efficient data transport and storage
High-reliability execution backend
Trusted visualization libraries
Lightweight UI prototyping
People often assume AI fails because the model isn’t good enough. But in practice, it fails when everything around the model can’t scale.
We helped this team go from potential to production—from fragile demos to robust deployments—without rewriting their vision.
They built intelligence.
We built the scaffolding.
You need someone who’s solved this problem before.
Let’s talk.
Adam Lewis Senior Software Engineer A coworker of mine built a Go CLI for the Harvest time-tracking API. It’s a solid tool, but every time I wanted to use it
Share Your Feedback