SandboxAQ has announced an expansion of its Switch AI offering, giving companies greater choice and control as they adopt and scale AI agents. Building on its open-source Switch Rooms offering, which brings people and AI agents together with shared, company-controlled context, SandboxAQ is introducing Switch Trust and Switch ROI to help organisations evaluate agent risk, measure business value and optimise performance across the agents, models and providers they choose.
SandboxAQ launched Switch in August 2026 as a Commons Clause Apache 2.0 open-source offering built around the premise that companies should have the freedom to choose and change their AI agents, models and providers without losing control of their context. Switch Rooms brings people and agents together across Slack, Microsoft Teams, Discord and other workplace tools, enabling teams to work with the agents they choose while retaining ownership of their shared context, history and knowledge. As AI technology evolves, organisations can adopt new agents and models without starting over or becoming dependent on a single provider.
Building an accountable AI workforce
The first generation of enterprise AI largely measured adoption through seats, usage and token consumption. AI agents represent a fundamental shift: they can make decisions, take actions and work alongside people, creating a new set of expectations for oversight and accountability.
Mount Sinai West’s MSW Lab is already working with Switch Trust to build G Board, a framework that enables physicians to create clinical applications using AI agents.
“In the MSW Lab, physicians build real healthcare software with AI agents, so we have to prove those agents behave safely,” said Dr Jeffrey Bander, Chief of Cardiology at Mount Sinai West and Associate Professor of Cardiology and Artificial Intelligence at the Icahn School of Medicine at Mount Sinai. “Switch AI gives us a way to build that trust into how we work, so our physicians can focus on solving clinical problems without having to become AI security experts.”
The new products extend that foundation with the controls enterprises need to move agents from experimentation into everyday work.
Switch Trust: Make every agent safe to put on real work by finding, testing and guarding every agent your company runs. It works in four steps:
- Discover: Inventories the agents, models, tools, MCP servers and dependencies running across connected environments and code repositories, including agents security teams may not know exist. It records ownership, source, dependencies, findings and severity so teams have a live picture of their agent footprint.
- Scan: Analyses agents in code and the cloud for security vulnerabilities. It combines static analysis with AI-assisted triage to filter out false positives, assigns severity-score findings, maps supported issues to OWASP agentic security categories and recommends fixes.
- Evaluate: Connects directly to an agent and attacks it with adversarial and functional probes, including prompt injection, jailbreak, data-leakage and instruction-adherence tests. Evaluate is extensible, so teams can bring their own evaluations alongside Switch’s built-in tests. Each attack gets a pass/fail result that rolls into an overall health score.
- Monitor: Protects agents in production. Teams can onboard an agent through the SDK, which captures what agents do, including prompts, responses, model calls, latency, sessions and policy events. Agents running in Switch Rooms are protected by guardrail policies automatically. Policies detect prompt injection, secrets, PII, toxicity and jailbreaks in live interactions, and Trust can block, redact, alert on or log unsafe behaviour while preserving evidence of what happened.
Switch ROI: Shows what every agent costs and delivers, with specific recommendations to lower cost and improve performance. It offers the following key features:
- Measure: Tracks spend and performance at the individual agent, session and turn level with budgets, burn rates and alerts so teams catch overspend early. Connects that spend to real outcomes, such as tickets resolved, pull requests accelerated or tasks automated, rather than infrastructure metrics like tokens or API calls.
- Prove: Calculates the value each agent produces against what it costs. Switch establishes the human baseline, tracks agents’ spend automatically and measures metrics like cost per outcome and net ROI so teams can show whether an agent is earning its keep.
- Optimise: Finds where spend is leaking and decides what to fix, reroute or trim. Switch ranks recommendations for each agent, each with an estimated monthly savings range and confidence level. They cover poor model selection, tools that are offered but never used, context that is resent instead of cached, retries and loops that repeat work, and output that runs longer than the task needs. Teams can tackle the biggest leaks first and lower cost over time.
The Switch platform brings these capabilities together to address a growing problem for enterprises: AI agents are multiplying faster than companies can determine whether they are safe, effective or worth the investment. Gartner predicts that more than 40% of Agentic AI projects will be cancelled by the end of 2027, while only 7% of leaders say they have established a return on their AI investments.
“More agents doesn’t automatically mean more value. We’ve seen this across many users: in cases where we thought agents could help avoid a net-new hire, the AI spend quickly became more expensive than hiring a full-time person,” said Mohammed Aboul-Magd, GM of Switch AI at SandboxAQ. “That’s why ROI can’t just mean ‘we adopted AI.’ You need to know what every agent costs, what it delivers and how to make it better over time. That’s the problem Switch ROI is built to solve.”
“We’re entering a world where companies will have hundreds, potentially thousands, of AI agents working alongside their employees. But you wouldn’t hire a thousand people without knowing their qualifications, what they’ll cost or whether they’re delivering results,” said Jack Hidary, CEO of SandboxAQ. “The same discipline needs to apply to AI agents. And as technology advances, companies must be able to switch to better agents and models without losing the knowledge they’ve built. The intelligence belongs inside the company, not the AI provider.”
