When Agents Vanish, Enterprise Work Must Persist: RHEOVELA’s Durable Process Model

Agents are inherently short‑lived: a model session may last only minutes, processes can exit, networks may drop, workers can crash, and models or tools may be upgraded while a task is in progress. In contrast, enterprise activities such as procurement, hiring, auditing, contracts, engineering changes, and customer issues often span hours, days, or months.

This creates a structural contradiction in agent systems: the executor is temporary, but corporate responsibility must endure . Storing long‑term tasks solely in an agent’s dialogue or memory leads to three possible outcomes after a system interruption: loss of work, duplicate actions, or an inability to determine the current task state.

RHEOVELA positions itself as a Dynamic Process & Durable Workflow Platform. It materializes governed capability plans into persistent Process Instances, allowing long‑running work to be recovered, approved, migrated, and audited.

The platform uses event sourcing to record workflow state. The current state of a Process Instance is deterministically folded from the event stream, so even if a process crashes, the system can reconstruct the identical state as long as the events remain.

RHEOVELA also employs an atomic, idempotent command pipeline to handle restarts, duplicate requests, and out‑of‑order messages. Side effects on external business systems are not automatically repeated when messages are resent.

On the executor model, RHEOVELA introduces a unified Work Item that represents humans, agents, traditional services, and robots. Different executors interact with the workflow through the same claim, lease, heartbeat, complete, or fail interfaces.

Claims, leases, and fencing tokens prevent two workers from processing the same task simultaneously. For example, if an agent comes back online after its lease expires, the old fencing token cannot overwrite the state of the currently valid executor.

RHEOVELA covers real‑world enterprise workflow requirements, including:

Human approvals and long waiting periods

Compensating actions after failures

Instance migration after workflow definition upgrades

Re‑planning due to environmental changes

Sub‑processes, pauses, and resumptions

Edge‑node synchronization

Checkpoints and disaster recovery

Legal hold and high‑availability locks

The platform records a complete evidence chain: who a task was assigned to, who executed it, under what authorization, what external impacts were produced, and how it was finally completed.

Traditional BPM tools often suffer from overly static process models that cannot absorb dynamically generated plans from agents, while pure agent loops lack long‑term state, reliable recovery, and process accountability. RHEOVELA sits between these extremes: it accepts governed dynamic plans but translates them into long‑running process instances with deterministic execution semantics.

The open repository provides event, command, and workflow contracts, OpenAPI specifications, and worker SDKs for Go, Python, TypeScript, Java, and Rust.

In this architecture, agents decide the next step, while RHEOVELA guarantees that the step does not disappear from the enterprise world even if the executor vanishes.

Project repository: https://github.com/axisrobo/rheovela-open