NOTE: Demo visuals include blurred data or synthetic placeholders to protect customer privacy.

At Abnormal, debugging rarely starts in one place. Engineers bounce between Airflow runs, Databricks tables, files in Google Drive, and context buried in Slack threads. Agents could help, but only if they can actually reach those sources without getting stuck on setup and authentication steps.

Three frictions showed up repeatedly:

1. Tooling didn’t scale: defining many tools via MCP consumed too much of the context window and became harder to maintain over time.

2. Retrieval was inconsistent: agents would attempt many tool calls, “spin” for long stretches, and still fail to pull the needed data.

3. Real-world access was hard: even a simple request like “check this Airflow DAG” required environment resolution, auth handling, and source-specific steps the agent didn’t reliably execute.

Ivan’s solution was to meet the agents where they work best: explicit, discoverable commands with instructions the agent can follow. He built a set of CLIs for the most commonly used Abnormal data source engineers, then paired them with markdown guides and a centralized discovery pattern so agents can quickly learn what exists and how to use it.

Core capabilities include:

• CLIs that encapsulate auth, environment, and source-specific setup for common engineering data systems.

• Markdown instructions that teach the agent how and when to use each CLI.

• Central discoverability via a shared Claude.md entry point so agents can find tools without bloating prompts.

• Cross-source workflows, where the agent can correlate signals across Airflow, dbt lineage, and Databricks state.

As Ivan put it, “the solution for how we now scale is through these command line interfaces, or what I'll reference as CLIs.”

In a demo, Ivan prompted an agent to investigate a failing dbt DAG, identify the cause of the failure, trace downstream models, and check whether stale data was being updated.

Instead of manual UI hopping, the agent used CLIs (including Airflow, dbt lineage, and Databricks access) to assemble a coherent report from a single high-level request.

The immediate impact is less about flashy demos and more about day-to-day reliability. With CLIs handling the messy parts, agents spend fewer cycles guessing and more cycles retrieving the exact context engineers need.

What changed for teams:

• Engineers: fewer “come back in 10 minutes and nothing happened” sessions, and more end-to-end investigation support.

• Testing workflows: agents can pull Databricks data directly for tests, avoiding manual UI exports and CSV wrangling.

• Shared momentum: Ivan shipped 10–11 CLI tools in a couple weeks, with visible engagement from engineers reacting to new tool announcements.

Next, Ivan intends to expand the library of CLIs and keep tightening the markdown playbooks so more investigations, not just the happy paths, are fully agent-runnable.

Early users have been vocal that these integrations make the agents feel less like code generators and more like reliable partners during investigations, especially around Databricks workflows that used to be painful and manual.

That enthusiasm is also shifting behavior. Instead of private DMs, Ivan is nudging engineers to share feedback in the AI for R&D discussion channel so adoption patterns and best practices become visible, searchable, and easier for new teams to copy.