Agentic Analytics and Auto-pilots

Transcript

1. Agentic Analytics and Auto-pilots Pete Hampton Principal Engineer - AI/ML Dublin, Ireland - April/2026
2. ©2026 CLICKHOUSE INC., CONFIDENTIAL & PROPRIETARY 2 Wild agents appear Dec 2024 Internal • DWH Assistant • Support Assistant • LogHouse AI • CH Assist • … External • ClickHouse Console Agent • clickhouse.build • ClickStack Notebook AI • …
3. ©2026 CLICKHOUSE INC., CONFIDENTIAL & PROPRIETARY 3 Agentic coding • Agents write code • Optimise queries • Improve performance • Automate testing • Ship faster
4. ©2026 CLICKHOUSE INC., CONFIDENTIAL & PROPRIETARY 4 Agentic coding • Typing boilerplate, finishing stale PRs, resolving merge conflicts • Porting code between codebases (e.g. a $500, 36-hour agent session fixing a SQL parser) • Bug investigation, incident response, fixing flaky tests 700 PRs submitted in Jan–Feb alone) • Security research 100% of real bug bounty findings now use AI • Vibe-coding internal tools; prototyping features cheaply • Getting cross-team work unblocked by submitting agent-generated PRs directly
5. ©2026 CLICKHOUSE INC., CONFIDENTIAL & PROPRIETARY 5 Agentic coding • Level 1: Copy-pasting from ChatGPT • Level 2: CLI/IDE agents with hand-holding • Level 3: Fully autonomous multi-agent loops
6. ©2026 CLICKHOUSE INC., CONFIDENTIAL & PROPRIETARY 6 Adoption of agentic coding (~mid 2025
7. ©2026 CLICKHOUSE INC., CONFIDENTIAL & PROPRIETARY 7 Not just demoware… • Bad queries • Wrong joins • Misread schemas • Slow and expensive • Inconsistent results
8. ©2026 CLICKHOUSE INC., CONFIDENTIAL & PROPRIETARY 8 Challenges & Risks • Hallucinations and incorrect outputs • Misunderstanding of schemas and data context • Cost and performance unpredictability • Lack of observability into agent behaviour • Trust and reliability concerns
9. ©2026 CLICKHOUSE INC., CONFIDENTIAL & PROPRIETARY 9 What Good Looks Like • Accurate, reliable outputs • Real-time performance at scale • Transparent and observable agent behaviour • Guardrails and domain knowledge built-in • Continuous learning and improvement
10. ©2026 CLICKHOUSE INC., CONFIDENTIAL & PROPRIETARY 10 clickhouse.build +
11. ©2026 CLICKHOUSE INC., CONFIDENTIAL & PROPRIETARY 11 Taming agents 1. Natural language to Tool Calls 2. Own your prompts 3. Own your context window 4. Tools are just structured outputs 5. Unify execution state and business state 6. Launch/Pause/Resume with simple APIs 7. Contact humans with tool calls 8. Own your control Flow 9. Compact Errors into Context Window 10. Small, Focused agents 11. Trigger from anywhere, meet users where they are 12. Make your agent a stateless reducer See: https://github.com/humanlayer/12-factor-agents
12. ©2026 CLICKHOUSE INC., CONFIDENTIAL & PROPRIETARY 12
13. ©2026 CLICKHOUSE INC., CONFIDENTIAL & PROPRIETARY 13
14. ©2026 CLICKHOUSE INC., CONFIDENTIAL & PROPRIETARY 14
15. ©2026 CLICKHOUSE INC., CONFIDENTIAL & PROPRIETARY 15 TDD → Offline] Eval Driven Development
16. ©2025 CLICKHOUSE INC., CONFIDENTIAL & PROPRIETARY 16 Model Context Protocol MCP LLMs mostly rely on training data MCP links AI apps with external tools and private data EXAMPLES • Company DB queries • Local file access • Smart-home control • Live service data Fully managed remote MCP server • Turn-key experience, no infra to setup or manage • Built into ClickHouse Cloud • Leverage your data with external agents and MCP-compatible clients • Secured with OAuth for authentication • Bring your own MCP-compatible client ◦ Claude, Cursor, Windsurf, 25 others
17. ©2026 CLICKHOUSE INC., CONFIDENTIAL & PROPRIETARY 17 Agents are great users of ClickHouse • Databases are becoming agent-facing systems • Query layers are being abstracted away • Natural language replaces traditional interfaces
18. ©2026 CLICKHOUSE INC., CONFIDENTIAL & PROPRIETARY 18 The shift • Dashboards donʼt answer questions • Exploration is slow • Insights arrive too late • The system doesnʼt act Queries Results Queries Results
19. ©2026 CLICKHOUSE INC., CONFIDENTIAL & PROPRIETARY 19 Agentic analytics SQL queries Prompts Prompts Responses Responses
21. ©2026 CLICKHOUSE INC., CONFIDENTIAL & PROPRIETARY 21 Why Now? • AI can write ClickHouse SQL. • Language is now the interface • Data systems are finally fast enough • Tools are becoming standardised • The stack is ready But is it 100% reliable?
22. ©2026 CLICKHOUSE INC., CONFIDENTIAL & PROPRIETARY 22 Agent skills • Agents need guidance • Skills = embedded expertise • Better queries / answers • Fewer mistakes • From demo → production https://github.com/clickhouse/agent-skills
23. ©2026 CLICKHOUSE INC., CONFIDENTIAL & PROPRIETARY 23 Agentic Analytics • Agents talk to your data • They write and run queries • They iterate until correct • They return insights, not raw data • Humans are optional in the loop Itʼs not just querying data - itʼs a system that can think through a problem using your data.
24. ©2026 CLICKHOUSE INC., CONFIDENTIAL & PROPRIETARY 24 The agentic data stack • ClickHouse → fast, scalable analytics • LibreChat → where questions start • Agents → think and plan • Tools → execute SQL, APIs) • Langfuse → trust and control This isnʼt just AI on top of data—itʼs a full stack designed for agents to operate reliably. https://github.com/ClickHouse/agentic-data-stack
25. ©2026 CLICKHOUSE INC., CONFIDENTIAL & PROPRIETARY 25 From Analytics to Auto-Pilots • Always-on monitoring • Detect issues instantly • Explain whatʼs happening • Take action automatically • No human in the loop
26. ©2026 CLICKHOUSE INC., CONFIDENTIAL & PROPRIETARY 26 Autonomous QA Agent - keeping CI Green 🟢
27. ©2026 CLICKHOUSE INC., CONFIDENTIAL & PROPRIETARY 27
28. ©2026 CLICKHOUSE INC., CONFIDENTIAL & PROPRIETARY 28 Auto-Pilot Analytics • Always watching • Detects instantly • Explains why • Acts automatically (scheduled) • No dashboards needed
29. ©2026 CLICKHOUSE INC., CONFIDENTIAL & PROPRIETARY 29 Takeaways & thanks 🙏 • Dashboards → agents • Insights → actions • Systems → autonomous • Closed loops / Trust → everything + + + + whatever you want