The Agent Harness We Tuned for Medallion

Before agents touch your data, they need a harness

How many people on your team spend Monday finding out why Friday's numbers broke? Now hand that same pipeline to a crew of AI agents with no rules. You get the same breaks, only faster, and nobody can say which agent did what.

An agent harness closes that gap. Snowflake describes a harness as the operating system around an AI model:

- Tools: What the agent is allowed to call, and how risky each call is.

- Guardrails and policy: Which data it can touch and when a human has to approve.

- Tracing: A record of every plan, tool call, result, and approval.

In other words, a harness controls what each agent is allowed to do by setting rules.

We built that harness into WALT and tuned it for medallion. It plugs into the stack you already run, including your data catalog and governance tools. Your data stewards own the business definitions it enforces, in files they can edit and version.

This post is for the platform lead whose CDO dragged them to our booth. You have ten minutes and a healthy distrust of anything called "autonomous." Good. Keep it.

How does the agent harness earn its keep?

As explained in a previous article, the medallion is a design pattern. Raw data lands in bronze, gets matched and cleansed in silver, and becomes reporting-ready in gold.

Building the medallion isn’t the hard part. The pain lives between the layers. Most enterprises have spent plenty on data. The architecture around it is where things stall.

Every promotion is a judgment call: is this data good enough for the next layer? Today a human makes that call, often at 6 PM on a Friday, often with one eye on the clock.

It’s also time-consuming. A March 2026 Fivetran benchmark of 500 senior data leaders found large enterprises take nearly 13 hours to resolve a single pipeline issue. And 53% of engineering capacity goes to maintaining and troubleshooting pipelines.

To speed this up, you can point an AI agent at that problem, but without rules, you get wrong answers faster, with no trail to explain them.

That's exactly the job of the harness.

What the harness does at every layer

Promotion gates

No data moves up a layer until it clears a gate. A gate is a set of checks the data must pass before silver or gold will accept it.

WALT's agents run those checks on every promotion. The Transformer (data transformation agent) builds the silver and gold models, and the Operator watches their data quality, freshness, and anomalies.

A failed check stops the promotion. You see which check failed and why, before a dashboard shows the wrong number to someone with a bonus riding on it.

Test generation

Gates are only as good as their tests. And nobody's favorite Friday task is writing them.

WALT's autonomous data engineering agents introspect your data the way a senior engineer would. They profile columns, map relationships, and read the business logic buried in your dbt models and query logs. What they learn becomes the checks.

Lineage

Every gate decision ties back to lineage, and that lineage lives in WALT's data context graph (ReasonBase^TM). WALT builds it by reading your dbt models, GitHub code, and database logs, so it knows which marts, dashboards, and reports sit downstream of each table.

When something breaks in silver, "what does this affect?" already has an answer in the graph. Nobody has to rebuild the map by hand on Monday.

5 things WALT’s autonomous data engineering agents refuse to do

A good harness is mostly a list of no's. Ours is short, and none of it is negotiable.

1. Let an LLM write production SQL: The language model only works out what you're asking. A deterministic engine builds the SQL from your context graph. Same question, same answer, every time.

2. Move your data: WALT runs as a container in your own AWS or Azure account, on your own LLM API keys. Your data stays in your cloud.

3. Change a business definition behind your back: Metrics and relationships live in human-readable YAML. Your data stewards edit them, version them, and roll them back.

4. Act without a receipt: Every action the agents take is recorded in a notebook your engineers can inspect. Every answer cites the source data behind it.

5. Ship code to production without you: Code changes land in a dev branch first. A human on your team approves the merge.

WALT does the work of a data engineer. Your platform team keeps the keys, the judgment calls, and the final yes.

Meet the Operator, WALT’s Platform Reliability Agent

My friends say that when a problem and I walk into a room, only one of us walks out. We built the Platform Reliability Agent with the same temperament, minus the drama.

It's an autonomous data engineer that works inside the harness, around the clock. The job fits in one line: observe every run, fix what breaks, re-run, and log the evidence.

How a Friday break plays out in 5 steps

1. Observe: WALT’s Platform Reliability Agent watches every pipeline run, including the ones that finish at 7 PM after everyone has gone home.

2. Diagnose: When a run fails or a gate rejects data, it walks the lineage back to the cause. A renamed column upstream. A late file. A join that doubled your row count.

3. Fix: It applies the fixes the harness allows and sends anything bigger to your team for approval.

4. Re-run: It reruns the affected steps and pushes the data back through the gates.

5. Log the evidence: It writes down what broke, why, what it changed, and which checks passed afterward.

Monday morning: Your team reads a post-mortem instead of writing one. Your platform team stays, but the 13-hour incident hunt goes to the agent. Their Monday goes back to the work they were hired to do.

Is this another tool to add to our already bloated stack?

Fair question. WALT works with the Snowflake, Databricks, dbt, and Airflow you already pay for, with no lift and shift required.

Everything the agents build stays with you: the context graph, the tests, the lineage. Fire us like you'd fire your systems integrator, and you keep all of it.

What we can show today:

→ 95% query accuracy: Across enterprise datasets, benchmarked.

→ 40%+ less report creation time: At a $4B agency.

→ 8 weeks: From zero infrastructure to a production-ready platform.

Watch it work live next week in Orlando

Next week, the Platform Reliability Agent runs live on our screen at Gartner IT Symposium/Xpo 2026. Bring your skepticism. We'll bring the evidence log.

→ Where: Booth 627, Data & Analytics zone, Xpo hall, Walt Disney World Swan & Dolphin, Orlando.

→ When: 19 to 22 October 2026.

→ The talk: Session TH88, Thursday 22 October at 1:15 PM EDT, Stage 2. "Your AI Is Waiting on Your Data Team. Let Agents Do the Data Work."

Book a time with us at the booth

Sources

1. Data Pipeline Failures Cost Enterprises $3 Million Per Month, Fivetran Benchmark Finds (Fivetran, March 2026)

2. Medallion architecture (Databricks)

3. What Is an Agent Harness? The Operating Layer for Agentic AI (Snowflake)

4. WALT AI: Your AI Is Waiting on Your Data Team. Let Agents Do the Data Work. (Gartner IT Symposium/Xpo 2026 session)