Wrapping up the AI team’s conference season

September and October traditionally mark “conference season”. This year, it has been “AI conference season”. I’ll explain. My talk “Your Database Has No Idea Who’s Asking” got accepted for PGDay Lowlands (talk recording), EDB CTO Max Romanenko showed off some of the new features of AIDB v7.7 at PgPune (zooming in on Governance), which my manager Jigar Mody and I got to do in more detail at PGConf EU in Valencia just last week. 

I co-organized the AI practitioners’ LatentShift, where I presented the keynote, and during the last day of the PostgreSQL Europe Conference’s “community events”, several colleagues presented during the “Autonomous Postgres” sub event. 

Presentation mode 

SQL is the second most used programming language, and almost one in three developers now use AI tools or agents to work with it, according to the Stack Overflow Developer Survey 2026. PostgreSQL is the most used database in the same survey. Agents already talk to databases. The question is how to make them effective and safe. 

In October, Jigar (EDB’s Chief AI Officer) and I (Staff Machine Learning Engineer) took this question to two conferences. During our LatentShift.AI keynote, “Your Agent Needs A Home”, we showed that only 5% of enterprise AI projects reach production. Teams add separate systems for memory, context, and governance, and each one adds cost and risk. Our reply: agents already have a home, the database. 

At PGConf EU we presented “Agent Ready Postgres” (slides). We showed how aidb, a Postgres extension by EDB, makes Postgres the logical home for AI agents*: 

• No rebuild. Agents, tools, memory, and semantics run inside Postgres. 
• Memory that works. Agent Memory got 85.0% accuracy on the public LOCOMO benchmark. 
• Right context. The Semantic Knowledge Base gives agents only the tables and queries they need. 
• Governance in the engine. Postgres checks every agent query before it runs and records every decision. 

At the PGConf EU “Autonomous Postgres” session, about AI and automation in Postgres, I showed why text-to-SQL agents fail on real enterprise data (slides will be available here). The model is not the main problem. Enterprise schemas have cryptic table and column names, undocumented changes, and business terms that no model saw in training. Without this context, an agent picks the wrong table, misses a sensitive column, or makes up an answer. 

The semantic knowledge base adds the missing meaning layer to Postgres. It uses the descriptions in COMMENT ON, the joins that foreign keys already define, and trusted queries that the domain expert approves. When a user asks a question, the agent searches by meaning and gets only the relevant information that question needs. For common questions, the agent runs a golden business query and does not write new SQL. The result is SQL that is more accurate, cheaper to produce, and based on the business context that already lives in the database.

Rajesh Nagpal and Anil Kumar opened that Friday morning with “Autonomous Postgres: From Self-Managed Operations to AI-Ready Data Platform”. They described how Postgres moves from a database that manages its own operations to a platform that is ready for AI workloads. 

Matt “The Yonk” Yonkovit followed with “Postgres Needs an Agent Loop: What Autonomous Database Operations Actually Require”. He described what an agent loop must have before it can run database operations on its own. 

Combined, our three talks showed one direction: Postgres can operate itself and serve AI agents, when agents get the right context and clear limits. 

What’s next? 

I’m already making plans for the 2027 edition of LatentShift.AI, and the team tracks open Call for Papers for next year’s events. I look forward to meet and discuss with more people in the community! 

In case you're curious, I interviewed Jigar at LatentShift, and the video is up now. 

* Most of this is available today in aidb 7.7.