Can CoCo build a working AI agent in 60 mins? We tested it w/ 100 developers & 107M GitHub events.

Poll results, step-by-step videos, & free code included.

Last week at the Snowflake SVAI Hub in Menlo Park, we ran a hands-on workshop with my team at the TechEquity AI Forum.

The challenge: take a room of 100 developers: more than half of whom had never opened Snowflake; & get everyone to a working AI agent in less than 60 minutes. No pre-built demos. No shortcut. A blank account & a GitHub dataset.

First the 3-minute overview if you’d rather watch how/than read:

This video covers the session intro, the stack, and what we set out to build.

[embed]The session stack, and what we set out to build. Credit: Zach Martin, Solutions Consultant, Snowflake

Before we started building, I ran 2 quick polls.

The first one: what’s your primary AI coding tool?

54% said Claude Code or Claude. Another 21% said they use multiple tools. This wasn’t a room of people wondering if AI belongs in their workflow. They were already there. The question they came with was which AI — not whether.

The second poll surprised me more.

45% had never used Snowflake before tonight. Not evaluating it, not on a free trial — never opened it. 16% were active customers. 23% were evaluating. Nearly half were starting from zero.

They all finished with a working AI agent.

What we built

5 steps. Same pattern works on any dataset.

Here’s what it looks like with GitHub data:

1. Load the data: 107M real GitHub events (30 days of public activity) loaded from a public S3 bucket via COPY INTO. Four minutes. Runs in the background.
2. Orient CoCo to the schema: One prompt. CoCo describes the table, identifies WatchEvent as the star signal, and recommends the right columns for trending analysis.
3. Build the trending repos query: A view of the top AI/ML repos by stars gained in the last 30 days. Real names. Real signal.
4. Add AI_COMPLETE: Wrap the SQL results in a natural language call. The query stops returning rows and starts returning sentences.
5. Wire the agent: Cortex Search + Cortex Agent = GitTrend. A named AI that answers questions about trending GitHub repos in plain English. No SQL required.

CoCo wrote every SQL statement. The devs directed it.

[embed]Follow along and build it yourself; Video Credit: Zach Martin, Solutions Consultant, Snowflake

When it stopped feeling like a tutorial

Step 3 is when it got interesting. The trending repos query comes back and people start reading names like OpenClaw out loud > repos they recognize, things they’ve actually starred or followed. Nobody told the agent what to surface. It just ran against 107M real GitHub events from the last 30days and returned github projects that are being built right now.

That’s when the exercise becomes real. Not a demo dataset. Not a walkthrough. Their agent, their query, real & recent signal.

What 54% Claude Code tells you

More than half this room already uses the same models that power CoCo: Claude & GPT > w/ Snowflake’s live data context layered on top.

So the question isn’t model quality. The question is what the model knows before it starts.

Claude Code walks into your codebase blind. It doesn’t know your Snowflake schemas, your roles, your warehouses. It has to ask. That’s what most of those extra tool calls are — orientation work that CoCo skips because it already has live context.

There’s benchmark data behind that. On ADE-Bench: 43 real data engineering tasks: CoCo scored 72.1%, outperforming Claude Code @ **65.1%. It uses 51% fewer tokens and takes 8% fewer steps** to get the same job done. Not a faster model. That’s what happens when the tool already knows your environment before the first prompt.

That’s the whole bet: the tool w/ Context wins. Not the tool w/ the best brain.

When attendees ran the trending repos query in Step 3, OpenClaw appeared in the results.

Take it further: give your agent the ability to export files

Once GitTrend is running, you can extend it in about 10 mins. Add a Python stored procedure, update the agent spec, and now when you ask it to “generate an HTML dashboard of the top 10 trending repos and give me a download link” > it actually does that.

Watch it work.

[embed]Extend your agent to export files: HTML dashboard with a download link

Build it yourself

Everything you need is open source & free:

📂 Workshop repo: https://github.com/sfc-gh-rbachala/building-ai-agents-with-coco-workshop

🎟️ Free Snowflake trial: here

Same 5 steps work on your product telemetry, support tickets, sales data, internal docs — whatever you have. The GitHub dataset is just the example. The pattern is what matters.

If this was useful: ⭐ star the repo; takes 2secs & helps others find it.

Thanks for reading!

[embed]Richie Bachala (@richiebachala) on X Agents, Interoperability & CoCo @ @Snowflake | past @Yugabyte @SherwinWilliams @WeatherfordCorp @HitachiOracle @Oracle…x.com