Cortex Code: The Agentic Reckoning,  Or, How Snowflake Finally Gave Data Engineers a Break from Their Existential Crisis (Now With 300% More Technical Guts)

Martyn Rhisiart Jones, Madrid, Thursday 26th March 2026

Yesterday I attended Snowflake’s breakfast date in Madrid, and here are some of the great things I learned. So, without more ado…

Listen up, you glorious data martyrs, you noble sufferers who’ve spent years knee-deep in the festering swamp of undocumented ETL pipelines, chasing lineage graphs that resemble a deranged spider on acid after a three-day bender, and muttering dark incantations at 3 a.m. because some crusty Python script decided “customer churn” meant “every table that vaguely smells like a customer, plus that one VIEW nobody documented since 2019.” I stand before you today, your erudite, slightly unhinged technical prophet (with a heavy dose of stand-up bile and a side order of schema diagrams), to deliver the good news: Snowflake has unleashed Cortex Code, the AI coding agent that doesn’t just autocomplete your misery,  it inhales it, digests the entire governed data estate, and burps back production-grade, hallucination-free SQL, Python, and dbt YAML while respecting your PII tags and warehouse economics like a paranoid compliance officer on Red Bull. This isn’t your garden-variety Copilot having another existential meltdown over a missing import. This is Code Context incarnate, an agentic beast that has swallowed the Horizon Catalogue whole, metadata, lineage graphs, semantic layers, role-based access controls, Dynamic Table lag policies, and the soul-crushing reality of your credit burn rate.

Picture the horror show: You’re staring at a blank editor in Snowsight Workspaces, contemplating another manual join across seventeen tables while praying the compliance team doesn’t notice the accidental public stage. Then you whisper into the void (or just type in the right-hand Cortex Code panel): “Build me an incremental dbt model for daily truck profitability that respects downstream lag, adds not_null tests, and optimises for TARGET_LAG = DOWNSTREAM so I don’t bankrupt the warehouse.” Cortex Code doesn’t blink. It fires up its agentic reasoning loop, powered by frontier models like Anthropic’s Claude 4.5 and 4.6 (optimised specifically for adaptive coding and multi-step orchestration), with cross-region inference via Cortex AI so you’re never left waiting while the best silicon is stuck in another AWS zone. It consults the Horizon Catalogue in real time: spots every PII = TRUE tag, traces column-level lineage to ensure no downstream regulatory fireworks, pulls semantic descriptions, checks your current role and privileges, and formulates a multi-step plan. Then it generates the code, slaps it into a diff view highlighting insertions and deletions (green for “I made this better,” red for “you were about to commit a war crime”), and waits for your trembling human finger on the “Apply” button before it even thinks about executing. Human-in-the-loop? More like “human-still-in-charge-because-I’m-not-a-complete-idiot.” And if the query blows up? There’s a big friendly “Fix” button in the results grid that analyses the error, suggests a corrected version, and shows you another diff. No more Stack Overflow tabs at midnight.

Technologically, this is a masterclass in not being stupid. Cortex Code is a persistent, stateful, multi-tool agentic orchestrator living inside Snowflake Intelligence. It maintains session context across follow-up questions, automatically selects the right tools (catalogue lookup, web search via Brave API in preview, or your custom skills), and executes with zero-copy access to live metadata. Type @ in the message box to pop up a real-time search of tables, views, and schemas; pick one, and it injects the full object context inline. Need to vibe-code an entire pipeline? It scaffolds DBT projects: explores sources, builds staging models, converts facts to date-partitioned beauties that incrementally update, adds tests, generates documentation, and even optimises Dynamic Tables for real-time freshness without you having to memorise the 47 parameters that make them not suck. In Notebooks, it spins up Snowpark Python for ML (pandas, scikit-learn, even CNNs for MNIST if you’re feeling fancy), visualisations, or full EDA. The CLI, now generally available, slots straight into VS Code or Cursor, turning your local terminal into a Snowflake-native agentic shell. Drop an AGENTS.md file in the workspace root, and it becomes a persistent system prompt: your coding standards, tribal knowledge, “never use SELECT *” manifestos, all version-controlled and shared like sacred texts. Custom skills live in .snowflake/cortex/skills, can be uploaded as files or folders, invoked with /skillname, and suddenly the agent has bespoke superpowers. It even plays nice with v0 by Vercel for Snowpark Container Services apps, Airflow, Jira, GitHub via Model Context Protocol, and whatever else you throw at it. Security? Enterprise-grade by design: everything runs under your existing RBAC (you’ll need SNOWFLAKE.COPILOT_USER plus either CORTEX_USER or the beefier CORTEX_AGENT_USER role), no credentials stored, full audit logs, and data never leaves your account. Regulated industries can finally let the robot touch production without summoning the compliance kraken.

The advantages? Let’s separate the nerdy glee from the capitalist cackling, with actual technical meat this time.

Technological high-fives (the part that makes engineers weak at the knees):

• Context so deep it needs its own submarine: Generic agents read your repo and still hallucinate because they lack enterprise data semantics. Cortex Code reads the governed platform itself,  Horizon Catalogue metadata, object tags, full column-level lineage, semantic descriptions, compute semantics, and governance policies. Result? Suggestions that actually respect your environment instead of producing beautiful but operationally suicidal SQL.
• Proper agentic orchestration, not just autocomplete: Interprets intent, formulates multi-step plans, maintains state, reflects, and iterates. “Ask Cortex” turns natural language into executable workflows; “Explain” translates your legacy window-function spaghetti into plain business English (“This monstrosity is trying to calculate running totals but in the style of a drunk accountant who hates indexes”); inline AI suggestions predict SQL continuations based on query history, table schemas, and recent executions. All while you retain veto power via diffs.
• Extensibility that doesn’t require a PhD: AGENTS.md, personal skills directory, Brave web search (toggleable by ACCOUNTADMIN), cross-region model access. It’s open enough to migrate existing agentic workflows, closed enough to keep the compliance team happy.
• Performance and economics baked in: Optimises warehouses, suggests rightsizing for the five biggest credit hogs, writes incremental models that don’t treat your compute budget like confetti. Dynamic Table lag awareness means your real-time pipelines stop being expensive science projects.

Business advantages (the part that makes your CFO do a little involuntary dance):

• Productivity isn’t marketing fairy dust. Customers report slashing the gap from vague idea to governed production: LendingTree moves from exploration to personalised AI-driven workflows in days; United Rentals cuts Natural Language Query tuning cycles dramatically; Shelter Mutual keeps regulatory controls while actually shipping; WHOOP benchmarks agents against evaluation sets for measurable accuracy gains; dentsu translates requirements into solutions without workflow disruption. The net result? Data teams escape the 80% plumbing tax and start doing actual, valuable work. Time-to-value collapses like a cheap deck chair in a hurricane.
• Democratisation without the usual chaos: Domain experts describe intent in human words (“Make my churn model actually useful for once”); analysts self-serve complex joins; engineers stop being syntax janitors and start architecting the future. Everyone wins except the irreplaceable gatekeepers (sorry, not sorry).
• Cost discipline and velocity: The agent knows your economics. It optimises, incrementalises, and rightsizes automatically. Innovation cycles shorten. Competitive moats deepen. And because it’s secure by design inside the Snowflake ecosystem, even the most paranoid insurance or finance teams can play.

Of course, we must nod to the competition with the weary sigh of a man who’s seen one too many “agentic” demos that collapse under real enterprise weight. GitHub Copilot, Cursor, Claude Code, Amazon Q,  they’re brilliant at understanding your codebase. They parse Git history like literary critics on a deadline. But hand them enterprise data semantics, and they look like a toddler asked to explain quantum entanglement while blindfolded. “Sure, here’s some SQL… Hope it doesn’t leak PII or explode your budget at 2 a.m.!” Databricks has its own agentic efforts with Unity Catalogue and Mosaic AI,  respectable, no question,  but in head-to-head data-specific tasks, the native Horizon Catalogue depth, live lineage integration, and Snowflake-native optimisation (Dynamic Tables, Snowpark, dbt workflows) give Cortex Code the edge, often with fewer awkward silences and more actual governance. Open-source RAG-over-metadata setups or custom LangGraph orchestrations are intellectually stimulating, but they require the kind of heroic engineering that makes you question your life choices at 3 a.m. Snowflake just productised the whole thing, catalog-native context, agentic reasoning, human oversight, CLI extensibility,  and put it behind a chat window and a diff view. Cheeky bastards.

In the end, dear comrades of the late-night query marathons and the 47-tab lineage diagrams, here’s the punchline with the technical flourish: Snowflake Cortex Code isn’t another incremental feature. It’s the moment the data engineering lifecycle stops feeling like Sisyphus pushing a boulder uphill while blindfolded, hungover, and carrying a legacy Python script that only runs on Tuesdays. By embedding true Code Context,  that profound, Horizon Catalog-native, governance-first, lineage-and-tag-aware understanding of your entire data empire,  into a secure, extensible, model-agnostic agentic orchestrator (Claude 4.x powered, cross-region, skills-enabled, AGENTS.md-guided), Snowflake has turned “vibe coding” from a buzzword into a daily, production-safe reality. The boilerplate is dead. The endless context-switching is extinct. The future belongs to those who type their intent (“Optimise this monstrosity and make it incremental, you magnificent silicon bastard”), lean back, watch the diff view, click “Apply,” and sip coffee while pondering their next domain-specific innovation.

If you’re still hand-crafting SQL in 2026 like it’s 2018, you’re not a purist. You’re a masochist with an expensive compute habit and a very understanding spouse.

Go on. Open Snowsight. Click the little Cortex Code icon in the lower-right corner. Install the CLI. Create an AGENTS.md that says “No SELECT *, ever.” Type something ridiculous.

The revolution will be agentically, hilariously, and mercifully executed,  with diffs, RBAC, lineage awareness, and human approval, of course. Because even in utopia, someone still needs to click “Yes, I trust the robot… this time.” And the robot will wait patiently, because it knows better than to execute without you.

Many thanks to Snowflake (Spain), to you for reading, and keep in touch.