EP 45: Why AI Agents Break in Production: The Missing Harness in Your Data Stack

Summary

This Data Science With Sam episode has Sam interview Pradmesh Patil of Altimate AI about why production data agents fail even when their base models are strong. The discussion uses Snowflake Cortex Analyst, hallucinated tables, silent wrong joins, expensive warehouse queries, and long-running data tasks to argue that data agents need a domain-specific Agentic Data Engineering Harness. Its core synthesis is that reliable agentic data engineering depends on AI Data Readiness, Deterministic Data Agent Validation, Data Agent Governance, Data Agent Context Compaction, and a human role shift toward Data Engineer Agent Supervision.

Key Claims

• The opening example cites an independent evaluation of Snowflake Cortex Analyst where six in ten AI-generated queries were wrong while still compiling and running.
• Pradmesh Patil argues that production failures are often harness failures: the model lacks enough domain context, ground truth, validation, governance, tools, or execution infrastructure.
• For data work, a harness should expose table schemas, lineage, query results, query profiles, query plans, tool access, workflow skills, sandboxes, and validation environments.
• Silent SQL Failure is more dangerous than syntax failure because hallucinated tables or wrong joins can produce plausible results that mislead users.
• Deterministic Data Agent Validation should check outputs where deterministic logic is available instead of sending every correctness decision back into the LLM reasoning loop.
• Altimate Code is presented as an open-source harness for agentic data engineering, with source-scoped claims about more than one million downloads, thousands of users, and strong benchmark results.
• Data Agent Benchmarks such as ADE Bench and DAB are used to argue that harness quality can be measured and that a stronger harness can outperform reliance on a larger model alone.
• Data Agent Governance includes cost controls, permission limits, PII boundaries, and cross-tool policy layers above fragmented warehouse, pipeline, and BI systems.
• Data Agent Context Compaction is treated as a correctness issue because generic compaction can drop schema or lineage details needed later in a long data task.
• The episode expects data engineers and data scientists to write less SQL or fewer dbt models manually while spending more effort directing, validating, and scaling fleets of agents.

Key Quotes

“6 in 10 AI-generated queries were wrong” - the opening evaluation example attached to Snowflake Cortex Analyst.

“system prompt tells the model what to do” - Sam’s contrast between prompt instruction and harness-grounded truth.

“recipe for disaster” - Pradmesh’s warning about compacting away schema context in long data-agent tasks.

Connections

• Data Science With Sam, Sam (Data Science With Sam), Pradmesh Patil, Altimate AI, and Altimate Code - show, host, guest, company, and open-source project context.
• Snowflake, Snowflake Cortex Analyst, Silent SQL Failure, and AI Data Readiness - data-platform and wrong-query failure branch.
• Agentic Data Engineering Harness, Agent Harness, Agent Runtime Execution Layer, and Agentic Workflow - harness and execution-infrastructure context.
• Deterministic Data Agent Validation, AI Verification, Agent Reliability Verification, and Output Quality Gates - validation and correctness branch.
• Data Agent Governance, Enterprise Agent Governance, Agent Permission Boundaries, and Model Routing Cost Control - cost, permission, and policy-control branch.
• Data Agent Context Compaction, Model Context Protocol, and Context Engineering - context-delivery and compaction branch.
• Data Engineer Agent Supervision, Data Engineering For Data Science, MLOps, and Machine Learning Engineering - role-shift and production-data workflow context.
• Data Agent Benchmarks and UC Berkeley - benchmark context mentioned in the episode.

Contradictions

• No direct contradiction found.
• The episode reinforces existing AI Data Readiness, Agent Harness, and Enterprise Agent Governance claims by specializing them to production data agents, where SQL can be syntactically valid but semantically wrong.
• Benchmark rankings, download counts, error-rate figures, and the expensive-query example remain source-scoped because the episode summary does not provide independent methodology or audit detail.