SandboxAQ turns team chat into the shared workspace for AI agents with Switch

SandboxAQ released Switch on Thursday, software that converts channels in Slack, Microsoft Teams and Discord into shared rooms where employees and AI agents hold the same context, resources and history, with the code available on GitHub at no charge under an Apache 2.0 licence carrying the Commons Clause. The company says Quick Start and Team editions deploy within minutes in a standard enterprise environment, and that the architecture extends to larger deployments without a vendor contract attached.

The release addresses the coordination problem rather than the model layer, which is where enterprise agent programmes have been stalling. Organisations that spent the past two years building or buying agents now hold a scattered estate of them, each sitting behind its own interface, each requiring the same background to be supplied again before it can do anything useful. Switch treats the messaging channel as the place where that background already lives.

"Every team we talk to has capable agents trapped in separate tools, and a coordination tax eating the productivity those agents were supposed to deliver," said Mohammed Aboul-Magd, General Manager, SandboxAQ.

The chat channel has become the contested surface for enterprise agent deployment

Slack and Microsoft Teams already hold the institutional record of how work happens inside most large organisations, which makes them the obvious surface for agents that need continuity across tasks. Salesforce has been building Agentforce into Slack, and Microsoft has been extending Copilot agents through Teams, both assuming that whoever controls the channel controls the workflow around it. SandboxAQ has entered that contest from the opposite direction, arguing that the channel should stay neutral and the agents should be whatever the team already runs.

That approach carries commercial logic for buyers who have watched their AI spend consolidate under a small number of suppliers. Switch connects through MCP, APIs and adapters, and supports agents built with Claude Code, Google ADK, LangChain and OpenAI frameworks, so a single room can hold agents from competing providers working on the same task. For enterprises negotiating renewals with hyperscalers and SaaS incumbents, the ability to run cross-vendor agents inside a channel they already own changes the leverage in those conversations.

The capability SandboxAQ is selling is memory attached to the work rather than to the agent. Rooms hold their own context, participants and rules, and preserve the timeline as people and agents join, leave and hand off, which means the same agent can be pulled into multiple projects without carrying contamination between them. Aboul-Magd described the effect as knowledge that compounds "instead of being relearned with every task."

The use cases the company describes follow from that design. A security team can bring a specialised agent into a live incident with no briefing, because the room already contains the timeline of what has happened. A marketing team can staff a campaign room with research, copy and design agents that each build on the previous output, and a legal team can add drafting and review agents that hold the full history of a deal rather than the current document. SandboxAQ says its own engineering, design and marketing teams used Switch to build Switch, running the development and launch inside rooms throughout.

The licence sits in familiar territory for commercial open source releases

Switch is released under Apache 2.0 with the Commons Clause, a combination that permits use, modification and internal deployment across an organisation, while reserving the right to sell the software itself or to offer a service whose value derives substantially from its functionality. SandboxAQ says users can run the software inside their enterprise, use it personally, and release free products built with it, which covers most of what development teams will want. The Open Source Initiative classifies licences carrying that condition as source-available rather than open source, a distinction that has followed the clause since Redis Labs first adopted it in 2018.

For the majority of likely adopters the difference is administrative. Enterprises deploying Switch internally and developers extending it face no practical constraint, and the free tier removes the procurement cycle that usually precedes an agent pilot. Organisations running policies that admit only OSI-approved licences will want a short review before adoption, and SandboxAQ retains the option to build commercial services around the project as it matures. The arrangement has become common across AI infrastructure, where companies release working code openly and keep a route to monetisation available for later.

The release fills a gap in SandboxAQ's agent portfolio rather than opening a new one

Switch sits within Flint AI, SandboxAQ's portfolio of agent products, and extends a path that began with the free Flint CLI, which developers use to scan and evaluate individual agents locally. The progression runs from testing one agent on a laptop to operating many agents alongside colleagues in production channels, which gives the company a presence at both ends of the agent lifecycle and a route into enterprises through developers rather than through procurement.

For SandboxAQ, the release also broadens a commercial base that has rested on Large Quantitative Models applied to life sciences, financial services, cryptography and navigation. The company spun out of Alphabet in 2022, has raised roughly $950 million across five rounds from investors including T. Rowe Price, Google, Nvidia, BNP Paribas, In-Q-Tel, Eric Schmidt and Marc Benioff, and was valued at $5.75 billion following its Series E in April 2025. Agent infrastructure carries none of the scientific-computing specificity of that core business, and a free distribution model gives the company reach into engineering teams that would never encounter its quantitative products.

Whether Switch converts that reach into anything durable depends on adoption inside teams rather than approval inside boardrooms, which is the wager the pricing reflects. "If you can message a colleague, you can put an agent to work," Aboul-Magd added.