Guides, case studies, and research on letting AI agents act on production data, safely, inside your guardrails.
AWS is acquiring DuckLabs, the company behind DuckDB. The code and trademarks stay with an independent foundation under MIT. The maintainers' payroll moves to a hyperscaler anyway.
Alice, formerly ActiveFence, raised $140M as the last scaled independent in AI agent security. Invariant Labs, Protect AI, Prompt Security, Lakera, and Portkey all sold to platform vendors first.
OpenMetadata 2.0 went GA on August 24. The breaking changes that fail your build are the safe ones. Three changed defaults, sampled profiling, dropped cardinality, MCP on, never error.
OpenAI acqui-hired the team behind InstantDB, the open source real-time sync backend. Instant Cloud shuts down within a year, and the agent state layer just lost another neutral vendor.
SAPinsider's 2026 benchmark puts 55% of SAP customers live on S/4HANA and only 34% finished. The gap is the data: a target model that deleted the tables the legacy data lives in.
Three critical CVEs and two quiet fixes hit Apache Polaris credential vending in four months. When the catalog mints cloud credentials, table names and storage paths become security policy.
Google's new Governance Agent propagates descriptions, policy tags, and trust scores through column-level lineage, free. What warehouse-native active metadata leaves for catalogs.
Agentic memory files grow 226 percent and almost never shrink. A study of 247,694 instruction lifetimes explains why: once a rule's rationale is lost, safe deletion costs O(2^n).
AI-assisted tools and agents mint more unique query shapes than pg_stat_statements can hold. Eviction is silent, and the tuning evidence disappears exactly where agent traffic grows.
Ossie's definitions were easy for 50 vendors to sign. The foundational semantics spec pins down joins, fan-out, and errors. A counter-proposal wants a smaller standard: SQL with measures.
PostgreSQL 19 Beta 3 is out and GA is close. REPACK brings online table rebuilds into core, and pg_plan_advice turns planner decisions into reviewable text. Both are built for automated operators.
OpenMetadata's 2.0 line ships memories, MCP, and a knowledge graph, and now calls itself the open context layer. What the rebrand says about where the data catalog category is going.
Databricks raised $5B at a $190B valuation and Lakebase, its Postgres for AI agents, crossed a $100M run-rate. The first proven agent-database business ships with platform governance built in.
Agent guardrails check permissions when an action is requested. Two new papers show the check can be false by the time the effect commits. Databases have owned this race for decades.
Tencent shipped team agent memory with ACLs. A new paper says permissions must live in the memory graph itself. Databases already learned why filtering at read time is not enough.
Datapace signs its first partner in Eastern Europe: Soft Industry Alliance, technology partner to manufacturing, supply chain, and retail companies modernizing systems and operational data.
Two new benchmarks score LLMs on the whole database lifecycle, not just SQL generation. The headline finding: tuning a model for text-to-SQL degrades the other four phases of the job.
Two unpatched MCP flaws chain into one attack: a poisoned tool list or injected instructions marked public-cacheable, then re-served to every user behind a shared gateway.
An AI-coded Postgres rewrite passed all 46,066 regression tests. A fuzzer broke it within days. What each layer of Postgres-compatible actually proves, and which layer decides adoption.
Neon's Object Storage, Functions, and AI Gateway hit beta: one branch_id now forks your database, files, functions, and model access together. What copy-on-write isolates, and what no fork can carry.
Supabase is now a connector on Perplexity Computer: hosted agents can read and write production Postgres across runs. A year after the MCP lethal-trifecta incident, the write path is the product.
Google's database agent finds root cause in minutes and executes fixes with your approval. That approval is now the entire trust boundary, and what the screen shows decides if it is a real gate.
A $1B week priced AI agent identity, but every identity product stops at the connection string. The Postgres mechanics of per-agent, short-lived credentials.
Article 12 requires high-risk AI systems to log events automatically. What the text demands, what it leaves open, when the duty actually bites, and how a database audit trail answers it.
Open Semantic Interchange entered the Apache Incubator as Ossie in July 2026. What the spec carries, what it has no place for, and why the gap decides whether an agent can trust a definition.
Mem0, Zep and Letta store what your agent experienced. None of them knows which revenue column is canonical. Where memory ends and a context layer begins.
The short answer: Mem0 for user memory, Zep for temporal knowledge graphs, Letta for agent-managed context, LangMem on LangGraph. Plus the one memory type none of them covers.
Why outdated documentation slows incident recovery, what living documentation means, and how to keep documentation up to date without discipline.
The two-layer setup for safe AI agent access to production databases, with measured blast-radius numbers and a gated path for schema migrations.
A context layer gives AI agents what's there, what it means, and how it connects. The definition, the components, how it differs from memory and semantic layers, and how to build one.
The six clauses a production database access policy needs once AI agents hold credentials, with a copy-paste template, a filled-in example, and the Postgres commands that enforce it.
OLTP, OLAP and HTAP measured on the same 50 million rows: point lookups 0.03 ms vs 0.22 ms, an aggregation 46x faster columnar, and co-location pushing the worst write from 38 ms to 332 ms.
Databricks coined LTAP in June 2026 and declared HTAP a failure. What actually changed: the row-to-column copy moved from inside the engine down into the storage layer.
Postgres indexes do not shrink automatically. Bloat compounds quietly until every read pays for it. Here is how to spot it and fix it with REINDEX CONCURRENTLY.
A runbook for giving AI agents real schema context: introspect the Postgres catalogs, harvest COMMENT metadata, map foreign keys, and assemble one context document.
The field pairs are the smallest part. A mapping document that survives to go-live also carries transformation rules, defaults, dedup and merge logic, exclusions, and one named owner per row.
A working method to reverse engineer an undocumented database schema: inventory tables, recover missing relationships, decode columns, validate the map.
How to find the source of truth when a database has duplicate columns like amount, total_amount, and revenue_net: trace writes, trace reads, reconcile.
An AI database reliability engineer watches production databases, diagnoses issues, and proposes fixes under human review. The role and the stack.
Legacy data mapping is where ERP projects slip after go-live. Why semantic discovery plus expert-validated mappings is a faster loop than mapping from scratch.
The console shut down on July 31, 2026. The API survives, AWS moved you to Database Insights Standard by default, and on-demand analysis plus plan capture are now Advanced-only. What to check now.
Agentjacking proves coding agents can be hijacked via MCP in 85% of attempts. Here is what it means for agent security gateway design in 2026.
One webpage, zero credentials, host-level RCE: AutoJack shows exactly why AI agents need a dedicated security gateway before production.
AutoJack RCE, WitnessAI GA, and Databricks DAIS 2026: three signals in one week that made AI agent security gateways non-negotiable.
Context windows hit 1M tokens; persistence is still zero. How memory infrastructure for AI agents works: 4 memory types, 3 production architectures, retrieval at scale.
A factual breakdown of the July 2025 Replit AI database deletion and the control-plane checks that would have stopped each failure.
Read-only credentials block the fixes agents are for, and prompt guardrails leak. What works: classify every statement in the data path, gate risky writes behind approval, and log it all immutably.
Most human-in-the-loop advice stops at approve database changes. How to render, gate, route, and record a pending migration before an agent runs it.
AI audit trail requirements for agents on databases: why API call logs cannot reconstruct what changed, and what a tamper-evident action ledger records instead.
Railway had two Postgres migration outages in six weeks. Same cascade, different trigger. A close reading of both post-mortems and the one pre- merge check.
Yahoo S5, NAB, and UCR shaped the TSAD literature. Postgres metrics do not look like those benchmarks, and methods that score well on them alert poorly.
The 2017 self-driving DBMS vision promised autonomous tuning, indexing, and healing. What actually shipped by 2026, and where a human still signs off.
Eraser (VLDB 2024) named it, RIB extended it. The robust-fix problem is one question asked at every layer of the DB stack. Datapace is building around it.
D-Bot, λ-Tune, and ROMAS are dashboard copilots. For regressions caused by a merged commit, that is the wrong surface. The case for database context at the decision point.
lock_timeout is correct and necessary, but not enough. A PR-time lock-graph simulator catches the same class of migration before the merge.
An April 2026 post-mortem tested a CHECK constraint on 10,000 rows, then hit 84 million in production. Four hours locked. The failure mode: sampling bias.
A SELECT issued during an ALTER TABLE waits even when locks are compatible. Why Postgres queues reads behind DDL, with pg_locks proof and the fixes.
The two Postgres diagnostic extensions are complementary, not competitors. When each lies, how to join them via queryid, and the minimum production setup.
Four online DDL tools, four architectures: versioned views, shadow tables, 2PC, and binlog streaming. A neutral comparison with failure modes named.
Between 2015 and 2025 the DBA role dissolved at mid-size SaaS. Three of four responsibility buckets found new owners. PR-time schema review did not.
DBSherlock, iSQUAD, D-Bot, and RCRank each infer a DB-internal root cause. None names the commit. The gap between inference and the developer workflow.
Supabase ships solid default monitoring. Past a point, three gaps appear: no historical context, a reactive top-queries view, no link to user impact.
Most Postgres performance problems come from a short list of avoidable SQL patterns: unindexed filters, SELECT *, deep OFFSET, N+1, and casts in WHERE.
EXPLAIN ANALYZE plans one query at a time. N+1 storms are a hundred fast queries adding up to a slow page. The signal lives in pg_stat_statements.
A sequential scan on a hot table is both a performance problem and a cost problem. The fix for both is the same index, caught at PR time.
pg_stat_statements is the only unbiased view of where Postgres spends time. Setup, the two queries that cover most use, and how to read the output.
Postgres has three levels of EXPLAIN: the plan, the actual execution, and the real I/O cost. Most developers learn the first and miss the other two.
Bring a use case. We will show you what agents can do on your live data, inside your guardrails.
