What if your Snowflake AI didn’t just spit out SQL — it remembered your team’s quirks, enforced your governance rules, and ran your custom playbooks without you lifting a finger?
Snowflake Cortex Code — or CoCo, as insiders call it — hides this power behind a simple chat interface. Most users stick to first gear, asking basic questions. But dig deeper, and you’ll find a full-blown agent architecture that’s quietly rewriting data engineering.
And here’s the thing: this isn’t some half-baked chatbot. It’s a multi-layered system with planning loops, persistent memory via AGENTS.md, built-in skills from Snowflake’s domain experts, and slots for your own custom skills. Think of it as upgrading from a bicycle to a Ferrari with a team of mechanics tuning it to your specs.
How Does CoCo’s Agent Loop Actually Work?
CoCo isn’t a chatbot that generates text and hopes for the best. It’s an autonomous agent that plans, executes, validates, and self-corrects.
Picture this: you fire off a prompt like “Analyze sales trends and flag governance issues.” CoCo doesn’t wing it. It kicks off a five-step loop — interpret intent, craft a plan, call tools, check results, respond with artifacts. That diagram in the original docs? It’s gold. Shows parallel tool calls merging into coherent outputs.
This loop is what makes CoCo fundamentally different from a text completion model. It doesn’t just suggest code — it runs it, checks it, and fixes it.
But — and this is where skepticism kicks in — Snowflake’s not reinventing the wheel. This echoes early agent experiments like Auto-GPT, but bolted onto a data warehouse. The genius? Tools tailored for Snowflake: SQL execution, object search, file I/O, even docs lookup. One prompt can trigger a barrage — say, schema inspection plus access audits via the governance skill.
Tools define the playground. Custom skills? They just chain these tools your way.
Short version: mastery here means prompting like a conductor, not a soloist.
Why AGENTS.md Feels Like Cheating
Forget starting from scratch every chat. Drop an AGENTS.md file in your workspace root, and CoCo ingests it every time — your persistent brain dump.
Naming conventions? Got ‘em. SQL style guides? Enforced. Cost guardrails (don’t query that 10TB table willy-nilly)? Baked in. Environment prefs, like favoring certain warehouses? Done.
It’s layer one of three concentric circles. Core agent at the center — raw power. AGENTS.md shapes defaults. Then built-in skills add Snowflake-specific smarts: cost intel, data quality, dbt workflows, ML ops. These fire on intent match, no extra config.
Layer three? Custom skills via SKILL.md files. Your team’s runbooks — deployment checklists, incident response, schema audits. CoCo spots the need, loads the skill, inherits everything below it. Layers compose smoothly.
My unique take: this mirrors how IDEs evolved from text editors to workflow orchestrators (remember pre-VS Code days?). Snowflake’s betting agents lock teams in deeper than raw compute. Bold prediction — by 2026, 70% of Snowflake data eng will be agent oversight, not manual SQL slinging. Corporate hype? A bit. But the architecture delivers.
The Toolkit: Where the Real Magic (and Limits) Hide
CoCo’s arsenal: SQL gen and exec, object discovery, file ops, Snowflake docs search. Parallel execution means efficiency — “What’s in SALES schema and who accesses it?” hits multiple tools at once.
Boundaries matter. Custom skills map to these; no magic beyond. Want a data quality checklist? Script it as tool calls in SKILL.md: query for nulls, cross-check lineage, flag anomalies.
Context awareness seals it — workspace files, chat history, AGENTS.md all feed the loop. No more “who am I again?”
But look, Snowflake spins this as revolutionary. It’s evolutionary — solid agentic AI on steroids for data pros. Skepticism: adoption hinges on that Quickstart Git repo. If it’s clunky, hype fizzles.
A three-word truth: Test it yourself.
Building Your Own: From Generic to God-Tier Assistant
Custom skills turn CoCo into your CoCo. Encode workflows: onboarding sequences, audit gates, even ML experiment trackers.
How? SKILL.md files in subdirs. CoCo detects context, loads, executes. Inherits AGENTS.md prefs and built-ins. Say you’re in a ‘deploy’ folder — boom, deployment skill activates, runs gates via tool calls.
This shifts architecture from user-AI to team-AI. Data eng becomes playbook curation. Why? Humans excel at workflows; agents at execution.
Historical parallel: Think Excel macros in the ’90s. Clunky at first, then indispensable. CoCo skills could do the same for warehouses — but Snowflake controls the moat.
Why Does This Matter for Data Teams Right Now?
Data workflows are brittle — SQL tweaks, governance slip-ups, cost overruns. CoCo agents systematize it. Persistent memory via AGENTS.md means institutional knowledge sticks.
For devs: less boilerplate, more innovation. Governance folks: automated audits. Execs: cost controls without micromanaging.
Critique the PR spin: Snowflake calls it ‘unlocking’ — fair, but it’s also retention glue. Once your skills are in, porting out hurts.
Deep dive payoff: see CoCo not as chat toy, but architectural bet on agentic future.
🧬 Related Insights
- Read more: AWS’s FinOps Agent on Bedrock: Cost Savior or CDK Nightmare?
- Read more: Zuck’s $200M Poaches and Tent Datacenters: Meta’s All-In Superintelligence Sprint
Frequently Asked Questions
What is Snowflake Cortex Code’s AGENTS.md file?
It’s a markdown file at your workspace root that CoCo reads every session, setting defaults like SQL styles, naming, and cost rules — your team’s persistent instructions.
How do custom skills work in Cortex Code?
SKILL.md files encode workflows (e.g., audits, runbooks); CoCo auto-detects and runs them via tool calls, inheriting lower layers for full customization.
Does Snowflake CoCo replace data engineers?
No — it handles execution grunt work, freeing engineers for high-level orchestration and playbook design.
Word count: ~950.
