Cloud.ru moves EvoClaw into general availability
Cloud.ru has launched commercial EvoClaw with an SLA, agent isolation and observability. We examine confirmed functions and B2B pilot requirements.
Contents
On 2 July 2026, Cloud.ru announced that EvoClaw had entered general availability. The managed cloud service for OpenClaw and other AI agents is now offered under paid plans, with a stated SLA and round-the-clock support. The provider also describes isolated workspaces, Zero Trust policies, integration with monitoring and secret storage, and logging and tracing functions.
What the GA status changes
General availability marks a move from test access to a paid service with contractual support and availability terms. That distinction matters more than simply launching an agent in a chat: a B2B customer needs a managed lifecycle, observability and a clear incident-escalation path.
EvoClaw supports OpenClaw and a modified enterprise version called NemoClaw. Cloud.ru says NemoClaw runs an agent inside an isolated environment where network, file and resource access is restricted by predefined rules. The service is part of Evolution AI Agents and can use catalogues of skills, prompts and MCP servers.
Where confirmed facts end
The source confirms commercial status, plans, an SLA, support and the listed management controls. It does not disclose numerical availability targets, responsibility boundaries, service regions, log-retention periods or compensation for an SLA breach. Those properties must come from current documentation and the contract.
The claim that an agent can be launched “in minutes” refers to initial deployment, not production readiness of a business workflow. Connecting corporate systems, assigning permissions, testing tools and agreeing on a safe failure mode remain a separate implementation project.
Why agent isolation is critical
A conventional language model produces an answer. An agent may additionally invoke tools, read files, call APIs and execute a sequence of actions. An error in an instruction, data or external content can therefore affect real systems rather than only generated text.
A minimum access model should include:
- a separate workspace for each use case;
- network denial by default with an explicit destination allowlist;
- short-lived credentials carrying minimum privileges;
- separate secrets for development, testing and production;
- limits on steps, execution time and cost;
- human approval for irreversible operations;
- a complete log of tool calls and changes.
Cloud.ru’s stated Zero Trust policies and isolation provide a foundation, but the customer still needs to verify the actual rules. A security principle’s name does not replace an access matrix and test evidence.
A practical B2B pilot
A first project should use a reversible internal task: collecting diagnostic data, drafting an operational report or searching an internal runbook. The agent can recommend an action but should not alter a production system without approval.
The pilot should cover several metric groups:
- Quality: correctly completed tasks and human interventions.
- Security: attempts to exceed permitted tools, secret access and policy triggers.
- Reliability: time-outs, retries, stalled chains and context recovery.
- Observability: ability to connect the outcome with every step, request and tool response.
- Economics: cost per successful task, including model use, tools, logs and human review.
The evaluation set must include abnormal conditions as well as normal requests: faulty instructions, an unavailable API, a corrupt file, ambiguous data and an attempt to gain excessive permissions.
Questions to answer before contracting
The announcement leaves several operational questions open:
- which model versions are available in each region;
- where requests are processed and traces stored;
- whether logs can be exported to an external SIEM;
- how OpenClaw, NemoClaw, skills and MCP servers are updated;
- which actions are stopped automatically;
- whether agent state has a guaranteed rollback mechanism;
- what exactly the SLA and round-the-clock support cover;
- how workspace data is deleted after a project ends.
Without those answers, an organisation cannot assess fit with its threat model and data requirements.
Limits of agent automation
Isolation reduces the impact of an error but does not guarantee a correct decision. An agent may misread an instruction, trust malicious content from an external page or repeat an operation after a time-out. Critical tools should be idempotent, while irreversible actions should remain outside the autonomous boundary.
An external API failure must not stop the core process. Our backup and fault-tolerance guide covers recovery patterns. Organisations selecting a platform may also compare IaaS with an owned data centre.
Takeaway
EvoClaw’s move to general availability shows the Russian agent-platform market progressing from demonstrations to managed services. An SLA, isolation, secret storage and tracing are necessary B2B controls, but they do not make every use case safe. Production deployment still requires least privilege, reversible operations, human review, a customer-owned evaluation set and contractually confirmed service parameters.
Primary source: Cloud.ru: commercial launch of the EvoClaw service
