Overview

Ezra is an enterprise multi-agent platform for production AI. A fleet of agents — of any size, spawning and terminating dynamically based on what the operation needs — share federated memory, federated data access, and a versioned audit trail of every commitment any agent made.

It is built for the case where one organisation needs many specialised agents acting in parallel, under permission boundaries, on time-critical work — and where per-agent decision latency must stay flat as the fleet grows.

The two jobs

Ezra does two distinct things at enterprise scale.

Data federation (the mesh). Connect to fragmented enterprise sources. Enforce permissions via a policy engine that inherits the user's identity or the agent's scope. Translate intent into pushdown queries that execute at the storage layer and return typed, provenance-tagged summaries — never raw data dumps.

Context management (the runtime). Decide what each agent in the fleet sees on every call. Manage what persists, what compresses, and what gets evicted across three memory tiers. Catch belief contradictions across the fleet and reconcile them before the model is called. Replay any session at any historical moment — and branch into counterfactual futures.

Most tools solve one or the other, for a single agent. Ezra does both, for any number of agents, as one platform.

What makes it different

• Cross-agent belief reconciliation — when two agents disagree on the same topic, a two-pass detector catches it and one of four strategies resolves it, before the wrong action is taken.
• Branching replay — reconstruct any prior state, mutate it, spawn agents that never existed, run forward, and diff against reality.
• Federated pushdown query — MongoDB, Snowflake, BigQuery, Atlas Stream Processing, REST — the database does the math; the agent gets a typed summary.
• Per-agent permission inheritance — scope is enforced at the federation layer, not left to the model to self-censor.
• Flat per-agent latency — adding agents adds rows, not iterations (see Architecture).

When to use Ezra

Ezra fits any domain with a dynamic fleet of scoped agents over federated data: incident response across infra / security / comms, financial operations across treasury / compliance / research, supply chain across procurement / logistics / inventory, or the reference demo — Formula 1 race-weekend operations, where telemetry, strategy, aero, parts, and press agents spawn and terminate across a weekend and contradict each other in real time.

Where to go next

• Quickstart — install, wire the backends, run your first agent.
• Architecture — the dual-layer model, session graphs, and how latency stays flat.
• The 8-Step Router — what happens on every agent turn.