Data agents that keep blindly probing a schema through repeated exploratory queries could get a big boost from one thing: an ontology layer that evolves from actual evidence. Title: EvoOntology: A Self-Evolving Ontology Layer for Data Agents URL: 📝 Overview EvoOntology is a self-evolving ontology layer, implemented as an MCP server, for agents that work across heterogeneous data like tables, files, and databases. ❗ Problem it solves Agents have no prior knowledge of data structure and must repeatedly issue exploratory queries. Static semantic layers need manual upkeep and can't adapt from execution history. ⚙️ Methodology Only candidates verified by actual probe queries get committed into the ontology. From there, agent trajectories surface intervention candidates, which are paired against the current state and adopted only when they show a clear improvement. 📊 Results On DDR-Bench it improved by an average of +17.8 points across 6 backbones, and on BIRD by +8.6 points, beating prior work — while cutting dialogue turns from 14.6 to 8.4 and reducing tokens by about 20%. 🔬 Use cases Even where static semantic layers actually hurt performance for some models, EvoOntology improved consistently — making it a practical fit for real-world data agent infrastructure. #DataAgents# #LLM#

We’re making it easier to develop agents—with new data agents and tools: 1. Conversational Analytics in BigQuery (in preview) 2. Data Engineering Agent (GA) 3. Data Agent Kit (in preview) 4. Managed MCP Servers for Databases (GA) + more →

Box CEO @levie says AI has been “unequivocally” a net positive for software. “You look at infrastructure providers—the Cloudflares of the world—totally on fire, because agents need sandboxes, they need compute, they need network, they need gateways. Great business.” “For us, we have reaccelerated growth far past our internal plans because it turns out that enterprises need core systems to be able to manage their unstructured data.” “Agents need to be able to work with that unstructured data to make decisions or move information through a workflow—whether it’s all of your contracts, your research materials, marketing assets, or financial documents.” “Now imagine an enterprise with 10,000 employees. How do you ensure that those agents continue to go after the actual canonical records, the actual documents that are the real sources of truth and the authoritative versions of that data?” “That means you need platforms.” “And all of that is creating more value for these platforms.”

Systems of Record aren’t going anywhere but … but … they are also extremely expensive databases as well. Way too expensive for agents. The cost per record, per GB, per anything in most Systems of Record is 100x-1000x-10,000x more than Postgres, Databricks, etc. And agents create, consume, touch and manage massive amounts of data. Far more than humans. So while classic Systems of Record aren’t going away, more and more of the data agents activate and use will be outside of them. Systems of Record’s own databases are just way too expensive, slow, and complex for most agentic data.

⚡ AI Brief ⚡ Databricks has raised $5 billion at a $190 billion valuation. The company originally sought to raise just $1 billion, but received approximately $15 billion in investor interest before ultimately accepting $5 billion, making the round roughly 3x oversubscribed. Founded in 2013, Databricks is a leading enterprise data and AI platform. Its Lakehouse architecture helps organizations unify their data and provides the foundation for AI models and AI agents. The round was led by Coatue, with participation from Blackstone, MGX, T. Rowe Price, Sixth Street, Point72, TPG, Clearlake, BOND, and Premji Invest. Existing investors including a16z, Thrive Capital, and Dragoneer also increased their commitments. Databricks has now surpassed a $7 billion annual revenue run rate, growing more than 80% year over year. The company plans to use the new capital to expand its AI agent capabilities, betting that in the AI agent era, competitive advantage will come from controlling the data agents rely on—not just building the most powerful models. Source: Databricks, TechCrunch

the @aiDotEngineer World Fair always one of the best events every year to talk to builders at the frontier of Research, Agents, Evals, Systems, etc A few weeks ago I gave a talk on - Continually Improving Agents - building Agents to understand data from other Agents - & a walkthrough of some of our latest work on data agents & post-training experiments some fun takes: - Every Continual Learning company will be an Observability & Eval company (and vice versa) - Environments & Evals are the currency of agent improvement. Agents are literally following the behaviors encoded in Evals. The best way to make good evals is mining Production data at scale - A good recipe to own your intelligence is using a Harness Eng - PostTrain - Harness sandwich with open models - Model-Harness-Task fit! There is no universal model or universal harness. You can always build a better agent system by optimizing the model and harness for a given task if your team is looking to understand your data at scale, build environment/evals, or just improve your agents - reach out, hmu would love to work with you! 🚀

# Practical and Useful Patterns for OpenAI Agent SDK 🌍 Add guardrails to tool inputs and outputs for enhanced security! Check tool arguments and mask outputs to prevent sensitive data leaks and unauthorized operations. 📌 Title: Guardrails – Tool guardrails 🔗 URL: 🧩 Overview Tool guardrails enable security checks before tool execution (argument checking) and after execution (output checking). They prevent API key injection, mask sensitive data, and block specific operations at the tool level. Even in complex workflows using manager patterns, Handoffs, or delegation, you can apply fine-grained checks to individual tools. 🛠 Usage Import `Agent` and `function_tool` from `agents`. Decorate a tool function with `@function_tool` such as `search_api(query: str) -> str`, then configure the agent with `Agent(name="SecureAgent", tools=[search_api], tool_guardrails=[check_tool_args])` to attach tool-level guardrails. 🏗 Practical Patterns **Block API Key Injection (reject_content)** Check if tool arguments contain API keys starting with `sk-` and block tool execution when detected. Import `re` and `GuardrailFunctionOutput`. Define `reject_api_keys(context, agent, tool_call)` which checks ` str(tool_call.arguments))` to detect API key injection, returning `GuardrailFunctionOutput(output_info={"checked": "api_key_presence"}, tripwire_triggered=has_api_key)`. Attach it with `Agent(name="SecureAgent", tools=[search_api, call_external_service], tool_guardrails=[reject_api_keys])`. **Mask Sensitive Data in Tool Output** Automatically mask sensitive information (email addresses, phone numbers, etc.) in tool execution output. Define `mask_sensitive_output(context, agent, tool_call, tool_output)` which applies `re.sub(r'[\w.+-]+@[\w-]+\.[\w.]+', '[MASKED_EMAIL]', str(tool_output))` for emails and `re.sub(r'\d{3}-\d{4}-\d{4}', '[MASKED_PHONE]', masked)` for phone numbers. Return `GuardrailFunctionOutput(output_info={"masked": True}, tripwire_triggered=False, modified_output=masked)` to pass the sanitized output. Configure with `Agent(name="DataAgent", tools=[query_customer_db], tool_guardrails=[mask_sensitive_output])`. **Per-Tool Checks in Complex Workflows** Apply fine-grained guardrails to specific tools even in complex workflows combining manager patterns, Handoffs, and delegation. Define `check_delete_permission(context, agent, tool_call)` which checks ` == "delete_record"` and verifies `context.get("user_role", "viewer")` is in `["admin", "editor"]`, triggering the tripwire for unauthorized users. Non-delete tools return `tripwire_triggered=False`. Combine multiple guardrails with `Agent(name="Manager", tools=[query_db, update_record, delete_record], tool_guardrails=[check_delete_permission, reject_api_keys])`. 💡 Use Cases 🔑 Prevent API key/secret injection in tool arguments 🎭 Automatic PII masking in tool outputs 🚫 Block specific tool operations based on permissions 🔒 Security control in complex multi-agent workflows ⚠️ Considerations - Tool guardrails are invoked per tool execution — consider performance impact - Regex-based checks are not exhaustive — use multiple defense layers for critical security requirements - Masking may change the original data type — verify it doesn't affect downstream processing - When multiple tool guardrails are set, all execute sequentially ✨ Use tool guardrails for fine-grained control over agent tool operations and enhanced security! #OpenAIAgentSDK# #AIAgent#