Kova — Durable Execution Engine for AI Agents
A WAL-First agent engine built in Rust — auto-recovers from crashes, 3μs scheduling
What is Kova?
Kova is Lurus’s core AI Agent infrastructure — a high-performance, durable execution engine built in Rust. It solves how agents reliably run for long periods, recover their state after a crash, and coordinate complex workflows. Traditional frameworks (LangChain, CrewAI) run in memory, so state is lost the moment the process exits. Kova adopts a WAL (Write-Ahead Log)WALWrite-Ahead Log. Every state change is written to the log before it executes, ensuring recovery from the log after a crash.Learn more →-first architecture: every execution step is persisted as a record, so even after a crash it can recover precisely to the point of interruption — without re-calling the LLM, losing progress, or incurring extra cost.
Key metrics
docs/benchmark-report.md), 315K ops/s throughput, and zero external service dependencies.Why choose Kova
WAL-First durability, microsecond-scale scheduling, zero-dependency deployment, and four access methods.
WAL crash recovery
Write-ahead log + CRC32 checksum per step; replays from the breakpoint after a crash, without re-calling the LLM
3μs scheduling latency
FIFO full-pipeline Criterion benchmark of 3.17μs, 315K ops/s throughput
Zero external dependencies
No Redis / Postgres required; runs with just a local WAL file
Four access methods
Rust SDK / gRPC / REST / MCP, modularized across a 21-crate workspace
WAL-First durability
All state changes are written to the WAL before they execute; on a crash, state is replayed from the WAL:
Agent decision — the engine determines the next action
WAL write (CRC32) — persists a record + checksum to guard against corruption
Execution — actually invokes the tool / LLM
Completion confirmed — marks the step as committed; on a crash, unconfirmed steps are automatically replayed
CRC32 checksums guard against corruption; a power-of-2 ring buffer makes efficient use of storage; and a strict Buffer → Queue → Txn lock order guarantees freedom from deadlock.
Agent orchestration
| Mode | Description | Use cases |
|---|---|---|
| Single agent | Executes a task independently | Simple automation |
| Workflow | Multi-step ordered execution | Data pipelines, approval flows |
| Swarm | Multiple agents collaborating autonomously | Complex research, multi-role simulation |
Tool ecosystem and multi-model support
Built-in tools (file / HTTP / database / shell), MCPMCPModel Context Protocol — an open standard that lets AI agents call external tools and data sources in a uniform way.Learn more → (connect to any MCP-compatible tool service, see the integrations directory), A2AA2AThe Agent-to-Agent protocol, which lets multiple agents communicate directly, delegate tasks, and exchange information. (direct agent-to-agent communication and task delegation), and custom tools (extend via Rust or a REST API).
Access all mainstream LLMs through the Lurus API (DeepSeek for everyday use / GPT-4o for reasoning / Claude for long contexts / Gemini for multimodal), switching dynamically per task at runtime.
Architecture overview
REST/SDK/gRPC/MCP access · Kova Core scheduling · WAL durability and recovery.
Kova REST API (Axum: 35+ endpoints, WebSocket)
→ Kova Core: Agent Engine(单/多 Agent) + Workflow(有序编排)
→ WAL + Ring Buffer(持久化状态管理 CRC32)
子 crate: kova-llm / kova-tools / kova-mcpKova is a workspace of 21 Rust crates, 178,284 lines of code, and 1,565+ tests (loom concurrency / proptest / chaos) plus 4 fuzz targets. It is currently a v0.2.0 pre-release (heading toward 1.0.0-beta.1), with strict lints fully enabled (#[deny(clippy::unwrap_used, clippy::panic, missing_docs)]).
Where Kova fits
| Scenario | Kova’s advantage |
|---|---|
| Long-running agents | WAL durability with automatic recovery after a crash |
| Complex workflows | Multi-step orchestration, conditional branches, parallel execution |
| Multi-agent collaboration | Swarm mode with direct agent-to-agent communication |
| Enterprise deployment | Rust performance, low resource footprint, no GC pauses |
| MCP tool integration | Native support for the Model Context Protocol |
| Security-sensitive scenarios | Optional encryption (SM4/AES), WAL HMAC integrity verification |
Get started by role
Compared with other agent frameworks
| Capability | LangChain | CrewAI | AutoGen | Kova |
|---|---|---|---|---|
| Language | Python | Python | Python | Rust |
| State persistence | None (needs external) | None | None | WAL-First |
| Crash recovery | None | None | None | Automatic recovery |
| Performance | Medium | Medium | Medium | Very high |
| Memory efficiency | Low | Low | Low | Very high |
| MCP support | Third-party | None | None | Native |
| A2A protocol | None | None | None | Native |
| Encryption | None | None | None | SM4-GCM / ChaCha20 |
| Multi-protocol | None | None | None | Four access methods: Rust SDK / gRPC / REST / MCP |
| Deployment form | Python process | Python process | Python process | Single binary / container / embedded library |