Orchestration and Execution: How JONI Approaches the Agent Layer

The distance between a model that produces correct output and a system that completes a task has turned out to be larger than most deployments anticipated.

A Workday survey of 3,200 employees across North America, Europe and Asia found that while 85 percent reported AI saving them between one and seven hours a week, roughly 37 percent of that saved time was consumed correcting, clarifying or rewriting low-quality output. Only 14 percent said they consistently achieved net-positive outcomes, and the heaviest users lost the most, with highly engaged employees giving up an estimated 1.5 weeks a year to rework. Workday characterised the cause as structural rather than behavioural, noting that AI has largely been layered onto roles never redesigned to accommodate it.

The organisational picture is consistent. IBM's 2025 study of chief executives found roughly a quarter of AI initiatives had met their expected return, and Gartner has projected that more than 40 percent of agentic AI projects will be cancelled by 2027, citing cost, unclear business value, and what the firm terms agent washing, the relabelling of existing automation as agentic systems.

The failure modes are reasonably well understood. Multi-step reliability degrades multiplicatively, so a pipeline of seven steps each succeeding 90 percent of the time completes less than half the time. Context does not persist between sessions. And most systems described as agentic terminate at output generation, leaving the final actions, provisioning, publishing, transacting, to a human operator.

JONI is one of several platforms attempting to address the third of these directly, positioning itself as an orchestration and execution layer above the foundation models rather than as a model provider.

Architecture

The system allocates each user a persistent cloud runtime holding memory, files, integrations and scheduled tasks, which continues executing background work between sessions and hibernates after approximately fourteen days of inactivity. Compute-intensive work is provisioned on demand as ephemeral instances and released on completion. All heavy processing runs in isolated sandboxes, with each user environment separated from every other.

The hybrid arrangement is a cost decision. Persistent per-user infrastructure carries a materially higher cost of goods than a stateless inference product, and hibernation plus on-demand burst compute is what makes always-on operation economically viable at these price points.

Model access runs through a gateway abstraction rather than direct provider integration, which allows substitution between providers without application changes. The company describes this as both an availability hedge and a commercial one, since provider pricing and terms remain the largest external variable in its cost base.

Routing

Task routing is handled by the platform rather than exposed to the user. Each request is classified and dispatched to whichever connected model the system judges best suited, with new models added as they are released.

The company's argument for this is structural rather than technical: a platform without a proprietary model has no incentive to route toward any particular provider, whereas each laboratory has an incentive to route toward its own. Whether automated routing outperforms informed manual selection is an open empirical question, and one the platform is well positioned to answer, since it observes performance across providers on identical task classes.

The company has stated an intention to publish comparative model performance data on a recurring basis.

Execution Layer

The distinguishing claim is that the platform completes actions rather than terminating at generation. Reported capabilities include domain registration, hosting provisioning and deployment of live sites with backend services and database persistence; construction and management of advertising campaigns through platform marketing APIs; publication to social platforms through official APIs with credentialed OAuth connections; media generation including multi-scene video with reference-based identity consistency verification; and telephony and email operating from dedicated addresses and numbers.

Actions are classified by consequence. Routine operations execute directly; consequential operations, including any procurement or outbound third-party communication, require explicit user approval before execution. All actions are logged to an audit trail available to account administrators, with reversal windows and a termination control.

For long-running unattended work the company describes stall detection with automatic restart, heartbeat recovery of orphaned jobs following host restarts, and checkpointing to resume mid-pipeline. These are unglamorous engineering concerns and largely determine whether multi-hour autonomous execution is usable in practice.

Extensibility

A marketplace allows third-party developers to publish agents and skills for installation into user environments, with revenue shared in the publisher's favour. Organisational accounts retain administrative control over which agents may be installed.

The structural argument is a two-sided network effect, with published agents attracting users and a growing user base attracting developers, expanding platform capability without corresponding first-party development.

Commercial and Market Context

JONI is sold on a per-seat licence at $65 per seat per month, with usage credits purchased separately into a shared account pool. Model capacity is described as purchased in volume and passed through at or near cost, with margin taken on the licence rather than on inference, which the company presents as a transparency position relative to vendors reselling a single model behind a proprietary interface.

The stated target is organisations of roughly five to two hundred people. Larger and more technically sophisticated organisations are described as a later priority.

Analysts size the surrounding category separately from generative AI overall. Deloitte estimates the agentic AI market at approximately $9 billion in 2026 rising to between $35 billion and $45 billion by 2030, conditioning the upper bound on enterprises implementing agent orchestration effectively. Gartner forecasts 40 percent of enterprise applications embedding task-specific agents by the end of 2026, against under 5 percent a year earlier.

JONI is developed by Mezada Development and Software Ltd., an Israeli company, and is self-funded. It is available on the web and via the App Store and Google Play.

Assessment

The orchestration layer is crowded. Enterprise platforms including Portkey, Langdock and Kore.ai already provide multi-model access with governance controls, and the major laboratories are extending their own products toward task execution. Multi-model routing on its own is converging toward a baseline expectation rather than a differentiator.

The claim worth testing is the execution one. Systems that provision infrastructure, transact, and publish are a small subset of those described as agentic, and the operational surface they expose, credential management, spend authorisation, failure recovery, action reversibility, is substantially larger than that of a generation product. Whether the reliability engineering holds at scale is the question that determines outcomes in this category, and it is not one that can be settled from a specification.