In the current agentic coding universe, human developers are building increasingly capable AI agents; agents themselves are building sub-agents… and the sub-agents are rapidly optimising the very infrastructure they all run on. All of which means agents (and the developers originating them) need integrations and repeated context sharing. SandboxAQ has now announced Switch, open source software for building teams of people and AI agents in the collaboration tools companies already use.
One of those “potentially quite significant companies that you don’t know as well as the hyperscaler behemoths and usual suspects of the frontier model world”, SandboxAQ is a software company that works at the intersection of AI and quantum techniques with Large Quantitative Models (LQMs) to deliver advances in life sciences, financial services, navigation, and other sectors.
What are Large Quantitative Models?
As they sound, LQMs are defined by the company as AI systems trained on first-principles equations of physics, chemistry, biology, and mathematics to model and simulate complex, real-world numerical systems rather than process text.
In terms of working functionality, Switch turns “channels” in tools such as Slack, Microsoft Teams and Discord into “shared rooms” where people and agents work together with the same context, resources, and history.
Switch is vendor agnostic and open source, so teams can use agents and models across providers without locking their workflows, context, or collaboration environment into a single AI ecosystem. The organisation says that Switch works through rooms, or channels that form around the work itself.
Humans & agents working side by side
“AI agents are rapidly becoming part of how companies operate, but most still work in isolation from the people they are meant to support,” said Jack Hidary, founder and CEO, SandboxAQ. “The real productivity gains will come from humans and agents working side by side, with shared context and goals. With Switch, we’re making that possible through open source, giving developers the freedom to choose their agents, models, and frameworks.”
Engineering and security teams can pull specialised agents into a live incident with no briefing required, because the “room already holds the timeline” and so it knows what context is present.
For example, a marketing team might staff a campaign room with research, copy and design agents, with every asset building on the last. A legal team can add drafting and review agents that already have the deal’s full context, making each round of review faster.
As agents come and go, the context, history, and rules stay with the work, so the team’s output compounds instead of resetting with every hand-off.
Switch is part of Flint AI, SandboxAQ’s portfolio of AI agent products. It extends the path from testing a single agent locally with the free Flint CLI, which developers use to scan and evaluate individual agents, to putting people and their agents to work together in shared rooms.
Key capabilities available in Switch include the fact that (as stated) it is open source, which here means that Quick Start and Team editions deploy in minutes, with an extensible architecture for enterprise deployment. Other key features include:
- Bring your own agents: Connect existing agents built with Claude Code, Google ADK, LangChain, OpenAI, and other frameworks
- Keep context with the work: Rooms preserve context and history as people and agents join, leave, and hand off work.
- One agent, many rooms: Use the same agent across projects and teams, with each room carrying its own context, participants, and rules
“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. “Switch pulls agents into the channels where teams already work, where knowledge compounds instead of being relearned with every task. If you can message a colleague, you can put an agent to work.”
What should developers think next?
For developers, SandboxAQ’s Switch may address a significant pain point i.e. the need to try and uphold and manage the issue of context fragmentation across isolated AI agents. The company here says it has the ability to bring framework-agnostic agents into environments like Slack and Teams, which (if it operates effectively over the long term) could lower orchestration overhead and preserve operational history. But there will (as there always is at this kind of level) be a degree of “adoption friction” generated if we want software developers (who, after all, are creatures of habit) to change the way they are used to working and now manage complex multi-agent state persistence, rate limits, and security permissions within open chat channels.
We didn’t even mention governance yet (so now we have) which will be an issue with open source flexibility in the mix so that we can avoid “context rot” (the degradation or dilution of an AI agent’s working memory as irrelevantly noisy conversational history accumulates) and unauthorised actions.
Switch is available today on GitHub and developers can get started with the Quick Start, and explore the documentation.