Landing
Why Ezra
Agent Context Graph
Memory
Branch
Audit
Join Us

Ezra — Agentic Infrastructure for Multi-agent context orchestration

Frameworks made it easy to build an agent. Ezra makes it safe to run a fleet.

Run more than one agent against real enterprise data and four things break at once: agents can't share what they learn, two agents prove contradictory "facts" and nothing is noticed, every agent queries raw databases with no policy layer, and nobody can answer what did the agent believe, from what source, at what time?

Ezra is the state layer under the fleet — what Letta (formerly MemGPT) did for one agent's memory, done for an entire agentic fleet's shared state:

Shared three-tier memory — Redis (hot) / Qdrant (warm) / MongoDB Atlas (cold, with per-turn Atlas Vector Search recall), shared fleet-wide and permission-scoped per agent at the source.
A versioned belief store — every commitment append-only and attributed. When two agents contradict each other, a two-pass checker (embedding gate → NLI) catches it and one of four reconciliation strategies resolves it by trust.
Git for agent state — rewind, revert, and branching counterfactual replay ("what if we prompted something else for this agent here") with diffs.
A federated data mesh — NL intent → validated, allowlisted native queries pushed down to Snowflake, BigQuery, and MongoDB (incl. time-travel queries and Atlas Stream Processing via MCP). The model sees typed summaries with provenance, never raw rows.
One audit feed — every agent's spawn, fetch, denial, commit, contradiction, and rewind as a chronological event stream.

Built on Google Cloud + MongoDB: the fleet is real Google ADK agents on Gemini via Gemini Enterprise Agent Platform (formerly Vertex AI), deployed on GKE with keyless Workload Identity; Atlas is the cold tier, belief store, audit log, and vector index. The demo's contradiction is genuinely LLM-emergent — detected and reconciled live, not scripted.