What does it take for an AI agent to answer a real business question? In this video, Shao, co-founder of Datus, walks through a real-world business scenario to show the problems Datus + DOSI can solve. Answering a question takes more than access to tables: an agent needs to understand the business concepts, relationships and definitions behind it. Datus helps teams build that shared context from documents, warehouses and ETL scripts. DOSI turns compatible semantic models into executable SQL, so agents can reuse reviewed definitions and teams can inspect the calculations behind an answer. Watch Shao connect the business question to the context and calculations needed to answer it—and see how Datus + DOSI work together. Explore Datus and DOSI: datus.ai/?utm_source=linkedi… Read the announcement: datus.ai/blog/datus-dosi-exe…

The dashboard is dead. The change-request backlog remains. “Add a filter.” “One more breakdown.” “Can we get a weekly view?” Somehow, the obituary did not close the ticket queue. Dashboards still earn their place in recurring monitoring and signed-off reporting. But a one-off question does not always need another permanent page—and another round of manual layout work. Generate the view for the investigation. Keep the metric definition under review. Give a view an owner and a maintained home when the team starts relying on it repeatedly. This is where Datus can help: Datus Agent can turn a question into an HTML view with KPI cards, charts, tables and filters, using existing semantic metrics and dimensions when they fit the request. New ad-hoc calculations still need review; live filters need the appropriate backend and datasource. Retire the busywork. Keep the useful view. What request keeps coming back in your dashboard backlog? datus.ai/blog/is-the-dashboa… #DataAnalytics #BusinessIntelligence #AIAgents #SemanticLayer

Meet DOSI, the semantic engine within Datus that turns Apache Ossie models into executable business context for AI agents. Watch our explainer video to learn how DOSI brings shared business definitions to life as native SQL—giving agents context they can use and teams calculations they can inspect. Explore DOSI: datus.ai/?utm_source=x&utm_m…

Introducing DOSI, Datus’s semantic execution engine. AI agents need business context they can act on—and calculations your team can inspect. DOSI turns compatible Apache Ossie models into native SQL, bringing shared definitions to agents and applications through CLI, MCP, REST/Arrow and Python. Explore Datus and DOSI: datus.ai/?utm_source=x&utm_m…

The HTML file is not the data service. An interactive dashboard can look finished while its next filter request still depends on the machine that created it. Before sharing, decide what the recipient needs: • A static snapshot with embedded results. • A live local view with the required backend and datasource. • A separately deployed application with authentication, access enforcement and refresh responsibilities. Those are different deliverables. This limitation describes the documented local workflow; Which service will answer your recipient’s next filter request? Read more: datus.ai/blog/composable-ana…

🚀 DOSI 0.1.13 + Datus Agent 0.4.3 are here. “How many users were active on exactly three days?” To answer that, you need to count active days per user first, then count users by that result. That’s a Level of Detail (LOD) question—and the calculation level changes the answer. DOSI 0.1.13 lets you define that level once and reuse it. LODs can work across datasets, use their own time window, appear in detail queries, and align two facts on a shared axis. Datus Agent 0.4.3 now uses DOSI directly to discover and query metrics, run attribution, and validate models. Also new in Agent 0.4.3: /init can analyze SQL dependencies and joins to help build a project map. That’s a separate project-initialization improvement, not part of LOD. Read the announcement: datus.ai/blog/dosi-datus-age… DOSI 0.1.13 release notes: dosi.datus.ai/release-notes/… Datus Agent 0.4.3 release notes: docs.datus.ai/0.4/release_no…

Datus-agent 0.3.8 is here! :rocket: 1. 𝗦𝗲𝗺𝗮𝗻𝘁𝗶𝗰 𝗦𝗤𝗟 𝗛𝗶𝘀𝘁𝗼𝗿𝘆 𝗣𝗿𝗼𝗳𝗶𝗹𝗲𝗿 `gen_semantic_model` now mines your historical and success-story SQL — column usage, data distributions, date spans, join coverage, common filters — so the OSI and MetricFlow semantic models it authors are grounded in how your data is actually queried. OSI authoring also uses your datasource's real SQL dialect and can sync complete OSI documents into the Knowledge Base. 2. 𝗙𝘂𝗹𝗹-𝘁𝗲𝘅𝘁 𝗶𝗻𝗱𝗲𝘅 𝘀𝗲𝗮𝗿𝗰𝗵 (𝗲𝘅𝗽𝗲𝗿𝗶𝗺𝗲𝗻𝘁𝗮𝗹) Metadata search can now use LanceDB native full-text indexing — with Chinese n-gram support — as an option alongside the default vector search. 3. 𝗕𝗲𝘁𝘁𝗲𝗿 𝗽𝗲𝗿𝗺𝗶𝘀𝘀𝗶𝗼𝗻 𝗰𝗼𝗻𝘁𝗿𝗼𝗹 𝗳𝗼𝗿 𝗯𝗮𝘀𝗵 𝘁𝗼𝗼𝗹𝘀 `execute_command` is now `bash`, with real shell pipelines plus command-level allow/deny/ask rules, project allowlists, per-call timeouts, and process-group cleanup. 4. 𝗕𝗲𝘁𝘁𝗲𝗿 𝗱𝗼𝗰𝘀 𝘀𝗲𝗮𝗿𝗰𝗵 English and Chinese docs now build as separate sites, each with its own search index, version metadata, and canonical URLs. Plus more reliable semantic storage (full catalog/database/schema/table identities) and recoverable agent/gateway failures under the hood. Full release notes: docs.datus.ai/dev/release_no…

Datus Agent 0.3.7 is out. 📷 1. 𝗨𝗻𝗶𝗳𝗶𝗲𝗱 𝗪𝗲𝗯 𝗦𝗲𝗮𝗿𝗰𝗵 𝗮𝗻𝗱 𝗙𝗲𝘁𝗰𝗵 𝗧𝗼𝗼𝗹𝘀 New `web_tool.web_search` and `web_tool.web_fetch` let agents pull in live web content mid-task. The backend picks itself based on your model — Claude native search, OpenAI hosted search, or local Tavily — and results render the same across CLI, TUI, and API. 2. 𝗨𝗻𝗶𝗳𝗶𝗲𝗱 `execute_sql` 𝘄𝗶𝘁𝗵 𝗯𝗲𝘁𝘁𝗲𝗿 𝗽𝗲𝗿𝗺𝗶𝘀𝘀𝗶𝗼𝗻 𝗰𝗼𝗻𝘁𝗿𝗼𝗹 `read_query`, `execute_ddl`, and `execute_write` merge into a single `execute_sql` that classifies each statement: reads (SELECT/SHOW/EXPLAIN) run automatically, writes and DDL ask first. The DDL restriction is lifted too — `CREATE DATABASE`, `TRUNCATE`, and `MERGE` now work once confirmed. 3. 𝗕𝗲𝘁𝘁𝗲𝗿 `gen_semantic_model` 𝗳𝗼𝗿 𝗢𝗦𝗜 OSI semantic models now consolidate at the business-domain level, `describe_table` syncs and shows table-level semantic profiles, and `ask_metrics` gets stronger metric discovery, cross-table join control, and cumulative/rolling metric queries. 4. 𝗕𝗲𝘁𝘁𝗲𝗿 𝗽𝗲𝗿𝗺𝗶𝘀𝘀𝗶𝗼𝗻 𝗽𝗿𝗲𝘃𝗶𝗲𝘄 Tool arguments in confirmation prompts now render as markdown — SQL as highlighted code, nested structures as JSON — so you know exactly what you're approving. A focused release, but an important step toward more 𝗰𝗼𝗻𝘁𝗿𝗼𝗹𝗹𝗮𝗯𝗹𝗲, 𝗽𝗿𝗼𝗱𝘂𝗰𝘁𝗶𝗼𝗻-𝗿𝗲𝗮𝗱𝘆 data engineering agents. Full release notes: docs.datus.ai/0.3/release_no…

Datus-agent 0.3.5 is here! 📷 1. 𝗗𝗮𝘁𝗮 𝗔𝗰𝗰𝗲𝘀𝘀 𝗣𝗼𝗹𝗶𝗰𝘆 Enforce row- and column-level access before any read SQL runs. The REST API passes structured caller identity (user, role, team) via `X-Datus-Principal`, and a custom `DataAccessPolicyProvider` can rewrite or reject queries before execution — wire it to OPA or your own auth service, or run the built-in `NoAuthProvider` out of the box. 2. 𝗦𝘁𝗿𝗶𝗰𝘁 𝗢𝗦𝗜 𝘀𝗲𝗺𝗮𝗻𝘁𝗶𝗰 𝗮𝘂𝘁𝗵𝗼𝗿𝗶𝗻𝗴 (Experimental) `gen_semantic_model` and `gen_metrics` can now author in strict OSI mode: generate OSI-compliant YAML, validate and dry-run it through the adapter, and write the metrics back to your knowledge base so the KB and semantic layer stay in sync. 3. 𝗣𝗲𝗿𝗶𝗼𝗱-𝗼𝘃𝗲𝗿-𝗽𝗲𝗿𝗶𝗼𝗱 𝗺𝗲𝘁𝗿𝗶𝗰𝘀 `gen_metrics` builds `offset_window` derived metrics (day/week/month/quarter/year) straight from your LAG SQL, and `ask_metrics` automatically matches the right time grain — so current, previous-period, and delta values come back together, with no missing first period. 4. 𝗠𝗲𝘁𝗿𝗶𝗰 𝗽𝗿𝗲𝘃𝗶𝗲𝘄 𝗔𝗣𝗜 New endpoints let a metric editor dry-run compile a saved metric into runnable SQL and list its queryable dimensions before saving. Invalid dimensions return structured preflight feedback instead of a raw error. 5. 𝗟𝗶𝗴𝗵𝘁𝗲𝗿 𝗶𝗻𝗶𝘁, 𝗻𝗲𝘄 𝗯𝘂𝗶𝗹𝗱-𝗸𝗯 𝗰𝗼𝗺𝗺𝗮𝗻𝗱 `/init` is now a fast, no-prompt project scan that writes an `AGENTS.md` inventory plus file-based knowledge/memory — no expensive vector store. Heavy KB construction moves to the new `/build-kb` command, scoped to your active datasource. 6. 𝗡𝗼𝗻-𝗶𝗻𝘁𝗲𝗿𝗮𝗰𝘁𝗶𝘃𝗲 𝗽𝗹𝗮𝗻 𝗺𝗼𝗱𝗲 `datus -p --plan-mode` auto-confirms generated plans, unblocking print-mode and benchmark runs that used to hang waiting for input. Plus: cached system-prompt prefixes cut input-token cost on long sessions, permission locks are now scoped per session so concurrent runs no longer block each other, the Claude subscription path fires permission/compact/lifecycle hooks consistently, and a round of CLI and connection-state fixes — more reliable across the board. Full release notes: docs.datus.ai/0.3/release_no…

Datus Agent 0.3.4 is out 🚀 AskMetrics — a dedicated metric QA subagent for KPI, trend & attribution questions. It reaches for metric-specific tools instead of raw SQL, so answers land at the right grain. Plus: scoped agent memory + faster knowledge extraction. docs.datus.ai/0.3/release_no…

From Modern Data Stack to Agentic Data Stack: The Next Evolution The Modern Data Stack (MDS) — cloud warehouses, ELT pipelines, BI dashboards — gave us powerful building blocks designed for human analysts and engineers. But what happens when AI agents become the primary operators? The key shift: MDS (2020): Optimized for human point-and-click workflows Agentic Data Stack (2026): Built for autonomous agent execution with human oversight This isn't just about adding APIs or CLIs. True agent-readiness requires: ✅ Structured error codes + recovery suggestions (not just toast notifications) ✅ Granular versioning (data, metrics, code — not just backups) ✅ Machine-readable observability (structured logs, lineage) ✅ Reusable skills (agents learn & improve over time) ✅ Dry-run modes for safe iteration The real test? Can agents complete 5 production workflows with <30% human intervention? The Modern Data Stack gave us the foundation. The Agentic Data Stack makes those blocks operable by machines. Small teams can start today with Git-based versioning, structured logging, and CLI-first interfaces. 📖 Full FAQ: agenticdatastack.org/blog/ag… #DataEngineering #AI #AgenticDataStack #ModernDataStack #MLOps #DataPlatform

We spent 2 months watching AI agents try to build real data stacks. Not because agents aren't smart enough. Here is the bitter lesson: Because data infrastructure was never designed for them. Traditional data platforms optimize for: → Query speed → Storage cost → Pretty UI But agents don't use UI. They parse structured data and CLI. They don't read tutorials. They need machine-readable contracts. They don't call support. They need actionable error codes. The shift: A platform scoring 95/100 on traditional metrics can be nearly unusable for agents. We identified 6 principles that separate agent-ready systems from chatbot wrappers: ✓ 𝐖𝐞𝐥𝐥-𝐝𝐞𝐬𝐢𝐠𝐧𝐞𝐝 𝐂𝐋𝐈/𝐌𝐂𝐏 𝐢𝐧𝐭𝐞𝐫𝐟𝐚𝐜𝐞𝐬 ✓ 𝐏𝐫𝐨𝐠𝐫𝐞𝐬𝐬𝐢𝐯𝐞 𝐝𝐨𝐜𝐬 + 𝐫𝐞𝐮𝐬𝐚𝐛𝐥𝐞 𝐬𝐤𝐢𝐥𝐥𝐬 ✓ 𝐕𝐞𝐫𝐬𝐢𝐨𝐧𝐢𝐧𝐠 & 𝐫𝐨𝐥𝐥𝐛𝐚𝐜𝐤 (𝐋𝐋𝐌𝐬 𝐦𝐚𝐤𝐞 𝐦𝐢𝐬𝐭𝐚𝐤𝐞𝐬) ✓ 𝐋𝐢𝐠𝐡𝐭𝐰𝐞𝐢𝐠𝐡𝐭 𝐜𝐥𝐢𝐞𝐧𝐭 + 𝐬𝐞𝐫𝐯𝐞𝐫𝐥𝐞𝐬𝐬 𝐬𝐜𝐚𝐥𝐞 ✓ 𝐃𝐮𝐚𝐥-𝐦𝐨𝐝𝐞: 𝐫𝐞𝐚𝐥-𝐭𝐢𝐦𝐞 + 𝐛𝐚𝐭𝐜𝐡 ✓ 𝐒𝐭𝐚𝐛𝐢𝐥𝐢𝐭𝐲 𝐨𝐯𝐞𝐫 𝐩𝐞𝐫𝐟𝐨𝐫𝐦𝐚𝐧𝐜𝐞 We built an evaluation framework + tested 27 open-source tools. Plus an experiment: AgenticDataTown — AI agents building an entire data stack autonomously. Read the framework → agenticdatastack.org/blog/ag…

🚀 Introducing Agentic Data Stack A community exploring how modern data systems evolve to serve agents — and how agents reshape data engineering. ✨ Vendor-neutral 📊 Practice-driven 🏗️ Architecture-first 10 tracks covering the full stack. Join us: agenticdatastack.org #DataEngineering #AIAgents #DataStack

Watched @peak ji’s sharing on Context Engineering for AI Agents (lessons from @ManusAI ) — learned a lot. My biggest takeaway: context engineering is the clearest boundary between “app iteration” and “model iteration.” Most real gains don’t come from changing the model, but from how we manage context. I especially like framing context engineering along these dimensions: reduce / isolate / offload / retrieve / cache — a very practical way to reason about agent quality and scalability. At Datus, we’ll keep pushing context engineering for data agents, and I’ve listed our current work items here: 👉 github.com/Datus-ai/Datus-ag… Furthermore, compared to general agents, specialized-purpose agents have real “special sauce”: Clearer rewards & benchmarks → much stronger feedback loops Parallel workflows + selection (with consistency checks) → noticeably higher reliability Domain-specific context → two evolvable context trees that make context offloading more standard, scalable, and maintainable Context isn’t just input. It’s an engineering system. #ContextEngineering #AgenticAI #DataEngineeringAgent #DatusAI #LLM

🎄 Merry Christmas from Datus 🎄 AI agents don’t fail because they’re not smart. They fail because they lack context. That’s why we’re building Datus — an open-source data engineering agent for Contextual Data Engineering, turning metadata, SQL, metrics, and feedback into a living foundation for reliable data agents. 🙏 Huge thanks to our community, users, and contributors for building with us this year. ⭐ If you like the idea and want to support us, please give Datus a star on GitHub: 👉 github.com/Datus-ai/Datus-ag… Happy Holidays — let’s keep making data agents work in the real world. #MerryChristmas #HappyHolidays #DataEngineering #OpenSource #DataAgents #Datus

The plan of next version(0.2.4) is out: github.com/Datus-ai/Datus-ag… 1) Dashboard Copilot (Auto-generation) Turn a BI dashboard URL into a scoped “dashboard copilot” (subagent) — automatically. We’ll also push attribution analysis forward with new tools, and start a BI adapter track for @apachesuperset . 2) Better SQL generation via a feedback loop We’re improving semantic model generation with CoT + probe SQL + reference SQL, and adding evaluation-driven knowledge generation (extracting “hidden knowledge” from success stories). Plus: a semantic layer adapter that supports more metric specs, and better context-tree recall via improved search tool specs + default prompts. 3) Datus-chat + Datus-cli UX Export/download, better SQL result rendering + chart visualization, and recent action history in CLI. 4) New use cases Catalog integrations (starting with @apachepolaris ) + interactive SQL editing with IDE. 5) Ecosystem expansion ClickHouse adaptor is on the roadmap. hashtag#DatusAI hashtag#DataEngineeringAgent hashtag#ReleasePlan

𝐓𝐡𝐞 𝐟𝐮𝐭𝐮𝐫𝐞 𝐨𝐟 𝐝𝐚𝐭𝐚 𝐚𝐠𝐞𝐧𝐭𝐬 𝐢𝐬𝐧’𝐭 𝐨𝐧𝐞 𝐨𝐦𝐧𝐢𝐬𝐜𝐢𝐞𝐧𝐭 𝐀𝐈. It’s a layered system of subagents—each with scoped context and purpose-built intelligence. Through months of collaboration with customers, we've identified a clear architectural pattern in Datus: 𝐚 𝐭𝐰𝐨-𝐥𝐚𝐲𝐞𝐫 𝐬𝐮𝐛𝐚𝐠𝐞𝐧𝐭 𝐨𝐫𝐠𝐚𝐧𝐢𝐳𝐚𝐭𝐢𝐨𝐧. 🔹 Layer 1: Foundational Subagents These are reusable “skills”(Native tools or MCP) every data team needs: • 𝐆𝐞𝐧𝐒𝐐𝐋 — generate + refine SQL with structural awareness • 𝐆𝐞𝐧𝐒𝐞𝐦𝐚𝐧𝐭𝐢𝐜𝐌𝐨𝐝𝐞𝐥 — draft dimensions/measures from sql & tables • 𝐆𝐞𝐧𝐌𝐞𝐭𝐫𝐢𝐜𝐬 — derive metrics from SQL patterns & business rules • 𝐒𝐮𝐦𝐦𝐚𝐫𝐢𝐳𝐞𝐒𝐐𝐋 — Summarize, classify and tag from SQL and put them into data context tree. • 𝐂𝐚𝐭𝐚𝐥𝐨𝐠𝐏𝐞𝐫𝐦𝐢𝐬𝐬𝐢𝐨𝐧𝐢𝐧𝐠 — Consider about governance & access to create a reasonable RBAC framework ... 🔹 Layer 2: Customized Domain Subagents Once you add scoped context (curated tables, metrics, SQL reference, domain rules), those foundational subagents become domain copilots: • Marketing Analytics Copilot • Supply Chain Operations Copilot • Restaurant Performance Copilot • Game Economy Copilot • User Growth / Retention Copilot Refer to: lnkd.in/gihuypH6 hashtag#DatusAI hashtag#DataEngineeringAgent hashtag#Subagent