EngineLakeAccessStudioProofGet it →

← agentdb.run

Notes for investors

The category, the proof, the record.

Category

AgentDB is the system of record for AI that must explain itself — one engine replacing the stitched-together nine-system stack (database, vector store, graph DB, prompt registry, eval service, trace store, audit log, access policy, workflow engine), for industries where “we can’t reconstruct it” has stopped being an acceptable answer.

Why now

Agents moved into regulated workflows — loans, diagnoses, refunds — and explainability obligations are hardening around them, while MCP standardized the socket agents use to reach data. The infrastructure to answer for AI decisions, with receipts, did not exist as one system. That is the product.

Proven — Wave 5 + Clean Sweep · July 2026 (losses included)

AxisVerdictFigureVerified
Durable OLTP throughput (TPC-C)LEADING10.69× vs SQLite synchronous=FULL2026-07-04 · c52131a
Durable writes (crud-bench, 9 competitors)LEADING5.2× creates/s vs SurrealDB's flagship RocksDB backend — their tool, their workload2026-07-04 · c52131a
AS OF point reads (branch time-travel)LEADING~16× AS OF point-read p50 vs Dolt 2.1.10 — the only other database with branches + SQL AS OF2026-07-07 · 3a02f7e
Vector ANN query (SIFT-1M, real corpus)COMPETITIVE4.5× query p50 + QPS vs pgvector 0.8.42026-07-07 · 220f09a
Multi-modal atomicityUNIQUEone transaction relational + vector + full-text + graph + audit in one commit, one snapshot2026-07-01 · e56a61f
Agent-native governanceUNIQUEshipped eval-gated deploy, exact answer replay, public-key-verifiable audit2026-07-07 · e56a61f
Vector ANN query (SIFT-1M)BEHIND~1.3× behind query p50/QPS vs Qdrant 1.15.5 (Qdrant leads)2026-07-07 · 220f09a
OLAP aggregationBEHINDbehind columnar aggregation vs DuckDB (DuckDB leads)2026-07-07 · 5633312

Full conditions, caveats, and runnable commands per row: /claims. Benchmarks run in numbered waves; the 2026-07-12 cloud re-lock put 3-run error bars on the standing ratios. The next wave re-runs the full set — verdicts may change; the dates and SHAs never do.

The assets

Five named subsystems, each with a one-clause proof:

  • The lake — your whole history on object storage you own; a machine dies mid-write and the next one picks up exactly where it stopped (fenced takeover, RPO=0).
  • Multi-tenancy — per-tenant logs, GC, and checkpoints; opt-in intra-tenant sharding for hot tenants.
  • AgentDB Access — built-in identity: passkeys and MFA for humans, scoped keys for agents, a sealed credential vault where use ≠ read.
  • Native MCP — the socket agents already speak, in-process, through the same gates as SQL; shipped June 2026.
  • AgentSQL — it’s SQL, with six verbs SQL never had: branch, time-travel, replay, eval-gate, deploy, blend.

Composability is the posture throughout: your object store, your KMS, your models (nine providers today, or fully local), your agents over MCP — embed the engine in-process or run it as a server.

IP & licensing

The engine is proprietary (© Cingulum Technology). SDKs and the published bench harness carry their own permissive licenses. Clean provenance: one codebase, one owner.

Where it is

GA declared 2026-07-13 for the positioned self-hosted scope (at v0.1.0; current release v0.1.2); engine and durability proven — the crash suites run with real SIGKILL in the executable release gate on every merge; design-partner program open; no managed cloud — deployments are done with partners, on their infrastructure. Full technical depth: /architecture. The forwardable summary: /brief.

Team

Sam Ginsberg — the engine is the résumé: one person wrote the Rust engine, the language, the console, the identity plane, and the benchmark discipline that publishes its own losses. Working with AgentDB means working directly with the person who built it.

Contact

sam@ginsberg.au with the subject tag “AgentDB — investor contact”. We reply within two working days.