Laurentiu Piciu

Building the data layer for AI agents. AI & data engineer: LLM agents and ClickHouse in production.

[Laurentiu Piciu @lpiciu23]

@pmddomingos doing this already: when I build an agent for a narrow problem, I let Claude Code solve it first. The tools and steps it builds for itself become the starting harness of the specialised agent.

[Laurentiu Piciu @lpiciu23]

@AltinityDB Thanks, glad it resonated. I'm open to it. What format did you have in mind?

[Laurentiu Piciu @lpiciu23]

I led our data platform migration from BigQuery to ClickHouse on AWS. The bill went from $53k a month to $25k.

The part I didn't plan for: on BigQuery, our AI agents' queries alone would cost about $37k a year.

The numbers:

[Laurentiu Piciu @lpiciu23]

Claude Code is the best agent harness I've used.
Opus 5.5 is the most capable model.

My rule for building agents, in any framework: give the task to Claude Code first, let it solve by hand, then read its transcript.

The tools it writes for itself become your agent's harness.

[Codez @0xCodez]

Anthropic engineers just released the best 1-hour course on building agents with Opus 5.5:

“Opus 5.5 is the best model for building agentic systems and reverse engineering with it.”

• 00:00 - introduction to Claude Managed Agents
• 07:45 - building Opus 5.5 Brain + architecture
• 23:40 - building /skills, /loops, and memory graphs
• 54:00 - agent dreaming & orchestration by Opus 5.5
• 1:10:23 - reverse engineering with Claude agents

this 1-hour watch will replace a $500 agentic engineering course.

watch it today, then explore how to reverse engineer anything with Opus 5.5 in the article below.

[Laurentiu Piciu @lpiciu23]

@marcklingen In private-credit document pipelines the source PDFs are the sensitive part, so even masked traces end up carrying deal terms. Self-hosted tracing sitting next to the data stops being a nice-to-have pretty fast.

[Laurentiu Piciu @lpiciu23]

@JimMarous @Vericast My guess is the gap is less about appetite and more about data. Generative AI works on a blank page; predictive and agentic need clean, connected customer data underneath.

[Laurentiu Piciu @lpiciu23]

The "present" most enterprises are stuck in is their data: undocumented tables, metrics nobody agrees on, documents still locked in PDFs. Agents expose that faster than any BI project did. What do you see working first: fixing the data before deploying agents, or starting with narrow agents built to handle messy data?

[Laurentiu Piciu @lpiciu23]

@nyike In finance the trace has to go one level deeper than the agent's decision: which document, which page, which clause a number came from. A CFO can live with an agent being occasionally wrong, not with being unable to show where a figure came from.

[Laurentiu Piciu @lpiciu23]

@hwchase17 Memory + skills is the box I'd watch. For agents inside a company, most of the useful "memory" already exists: tables, metric definitions, dbt docs. Curious whether context hub ends up pulling from the warehouse / semantic layer, or stays mostly agent-written notes?

[Laurentiu Piciu @lpiciu23]

My takeaway: before you give agents your warehouse, check how it charges you per query.

If you run agents on a pay-per-scan warehouse, do you know what one question costs?

[Laurentiu Piciu @lpiciu23]

AI. In 30 days our Slack data agent and analysts' coding assistants ran about 25k queries and read about 490 TiB.

At BigQuery on-demand prices: about $3k a month, up every month.

On ClickHouse: nothing extra.