The pattern repeated across every engagement. We'd build a beautiful data platform — Snowflake, dbt, Airflow, the works. Six months later, a senior engineer was still spending two days a week translating Redshift stored procedures to Oracle for a migration. Or an underwriter was retyping numbers from PDF SOVs into Excel. Or an L1 oncall was hunting through six tabs to find the right runbook.

AI chat tools could obviously help. But they couldn't safely do the work — they couldn't read your filesystem, run your queries, trigger your DAGs, or send your emails. Every demo was "copy this prompt, paste it back here." That's not an agent. That's homework.

What was missing

We needed an agent that could:

• Run on the user's machine, with the user's credentials. No SaaS, no shared API keys, no data egress.
• Connect to live systems — Airflow, Oracle, Postgres, Microsoft 365, S3 — through tools, not chat tricks.
• Carry context across sessions through markdown playbooks teams could read and version.
• Build apps — scope, plan, tasks, mappings — that survive turnover and audit review.
• Prompt the user when about to do something risky. No silent shell execution.

The bet

We bet that the right shape was a CLI for engineers, a Desktop app for business users, a VS Code extension for developers, and (now) a browser-based InsightStudio for everyone else — all sharing one agent core and one bundle format. After 18 months of building it with insurance and FinServ customers, that bet is paying off — underwriters drafting policy comparison briefs in minutes instead of hours, data engineers shipping migration validators in a day instead of a sprint.

InsightWorker isn't trying to replace the senior engineer or the experienced underwriter. It's trying to give them back the eight hours a week they currently spend on work nobody loves doing.

What's next

The trial is available now — CLI, Desktop, VS Code, and the new --worker mode that joins an InsightStudio farm. All five LLM providers (Bedrock, Anthropic, OpenAI, Azure, custom GPU endpoints) and a growing library of domain skills baked in. Try it on a real app this week. Tell us what's missing.