Agent-Native Data Infrastructure

1. Why agent-native data infrastructure is a thing now

For thirty years databases optimised for two consumers: humans writing SQL in a console, and applications running parameterised queries. In 2026 the third consumer became dominant: AI agents that discover schemas at runtime, issue thousands of exploratory queries per task, spin up isolated workspaces for trial-and-error, and need fresh full-fidelity context -- not sampled, not rolled up. PingCAP reports 90% of new TiDB Cloud clusters in Q2 2026 are created by AI agents. That single number changes everything about how a database should be designed. Six 2026 launches independently codified the shift: Databricks Genie Code (autonomous data engineering agent, March 11), Confluent Streaming Agents (Flink + MCP + Real-time Context Engine), RisingWave v2.6 (native vector + HNSW + official MCP server with 100+ tools), CockroachDB managed MCP server + agent-ready ccloud CLI (March), ClickHouse clickhousectl beta with agent skills installer (April 9), and PingCAP TiDB positioning as 'the database for agentic workloads'. Six vendors, same six design principles, no coordination -- that is what tectonic shifts look like.

3. The 2026 reference architecture -- four layers

Layer 1 -- Transactional write layer (Postgres / MySQL / TiDB). The system of record. Your agent does not write here directly; humans, apps and downstream agents do. Layer 2 -- Streaming compute (RisingWave or Confluent Flink). CDC-ingests Layer 1, runs continuously-updated SQL materialised views over the stream, joins with reference data, computes embeddings inline. This is where the agent's working memory lives -- fresh by construction, no batch lag. Layer 3 -- Serving layer (Postgres protocol + MCP). The agent queries Layer 2 via standard SQL over the Postgres wire protocol, OR via the MCP server exposing the same data as tools. Same data, two surfaces -- code agents pick SQL, conversational agents pick MCP. Layer 4 -- Vector + semantic search. Lives inside the streaming layer (RisingWave native vector(n) + HNSW + cosine + L2), not as a separate Pinecone-shaped system. One less data plane to govern, one less consistency model to reason about, one less place for memory to go stale. Mapping to the Day-44 Write-Aside pattern: Valkey/Redis is still your L1 hot cache (<5ms), but L2 changes -- instead of an async flush into pgvector, your L2 becomes a RisingWave or Confluent materialised view that is automatically maintained against your transactional store. The agent never sees stale memory, and you remove an entire class of write-aside race conditions. MemOS L3 sits on top as the unified API. This is the single biggest unlock in agent memory architecture this quarter.

4. Genie Code, Streaming Agents, and what they actually do

Databricks Genie Code (GA at FabCon 2026) is a fully agentic data engineering partner inside Lakeflow. It builds pipelines, debugs failures, ships dashboards, monitors AI models in the background, triages anomalies, autonomously analyses agent traces to fix hallucinations, tunes resource allocation before a human intervenes. On Databricks' real-world data science benchmark it more than doubles the success rate of the leading coding agent (77.1% vs 32.1%). Integrated with Unity Catalog so it enforces existing governance + access controls and understands business semantics + audit requirements -- Annex III

Confluent Streaming Agents (GA in Confluent Cloud for Apache Flink) is the event-driven counterpart. Agents are first-class Flink objects: they monitor real-time topics, reason via embedded LLM calls, act through MCP tool calls -- with every tool interaction logged into Kafka for replayability + auditability. The Real-time Context Engine continuously materialises enriched enterprise data sets into a fast in-memory cache and serves them via MCP, all fully managed. Event-driven by design -- agents communicate without rigid dependencies, collaborate with other agents, adapt to real-time inputs. This is the substrate the Day-49 Agent SRE will actually monitor in production. RisingWave v2.6 is the open-source third leg. Streaming SQL + CDC + incremental materialised views + native vector(n) + HNSW + an official MCP server (risingwavelabs/risingwave-mcp) exposing 100+ tools. openai_embedding() runs inside a SQL query or MV definition so embeddings stay fresh automatically. Apache 2.0, self-hostable -- the EU-AI-Act-safe default for agent memory when data residency matters.

5. Governance, EU AI Act, and the next 12 months

EU AI Act Annex III enforces Aug 2 (T-77 days). The audit checklist mandates reconstructable reasoning: what the agent did, why, with what data, in what order. A streaming substrate makes this near-trivial -- every materialised view update, every MCP tool call, every vector lookup is already a logged event with timestamp + SVID + nonce. Compare to a batch-ETL world where evidence is reconstructed retroactively from logs in three different systems, often inconsistently. The streaming agent stack converts an audit nightmare into a Kafka topic. Mapping back to Day 54 (SPIFFE + KYA): each agent SVID becomes the natural scope key for its streaming materialised views and MCP server namespaces. Per-agent isolation at the data plane, not just the app plane. Combined with Microsoft Agent 365 (control plane), AWS Bedrock AgentCore Payments (commerce) and Confluent / RisingWave / Genie Code (data), the 2026 enterprise agent operating model is now complete. Expectation for Day 56+: data-plane primitives become the locus of competitive differentiation as model performance plateaus and the moat shifts to context freshness + retrieval

MARKET SIGNAL

Anthropic in talks for a $30-50B raise at a ~$950B valuation; PwC announces firm-wide Claude deployment for client work; Anthropic + Gates Foundation form a $200M partnership. OpenAI grants EU access to GPT-5.5-Cyber in limited preview while Anthropic still withholds Mythos. US CAISI now pre-tests Microsoft, Google and xAI models before launch. Translation: data-plane and model-plane consolidation is happening at the same time -- pick your governance substrate before the regulator

Viral App of the Week

Open-source desktop AI agent (Rust + Tauri, MIT, github.com/tinyhumansai/openhuman, 5.6K+ GitHub stars in days, #1 GitHub Trending May 13-16). Inverts the playbook: instead of waiting for your first prompt, it pre-loads a 1B-token Memory Tree from 118+ services (Gmail, GitHub, Slack, Notion, Stripe, Calendar, Drive, Linear, Jira) via one-click OAuth. A 'subconscious loop' runs without user input -- reads your to-dos and recent memory, decides what to act on next. A desktop mascot can join Google Meet as a separate participant, transcribe into the Memory Tree, speak back. All local-first: SQLite + Markdown Obsidian wiki, no data egress. First credible consumer surface for the 'agent that knows you on day one' pattern.