Picture the most capable person you ever hired. Faster than anyone in the building, never tired, never bored, never irritated by the grind. And then imagine they refuse to use a single one of the team's tools. They don't come to standup. They never open the board. Finished work arrives as a wall of text someone else has to re-key into the systems by hand.
You would not keep that person. Not for lack of skill — the skill is obvious — but because they never became part of how the team actually runs.
That is the agent most companies have deployed. The capability is real. The presence isn't. You operate the agent from the outside — you bring it the work, you carry back the result. Useful, often genuinely so, and still in the hallway — never in the room.
Capability was never the bottleneck
We spent two years making agents capable. We spent almost none making them present. The agent that matters now doesn't need a higher IQ — it needs a chair at the table and the same rules of conduct as everyone else at the table. An agent that can't be told "not like that" mid-task, that can't be handed work and reviewed on it, isn't a teammate. It's output you have to go fetch.
An agent is a resource on the team
Management, stripped to its core, is a single act — putting the right task in front of the resource suited to it. Knowing what to delegate and what to keep. Knowing when to step in and when to stay out of the way. An agent is subject to exactly that discipline. It will be excellent at some things, adequate at others, and need a human close on the rest. So we treat it the way you treat any resource — it gets a role, it gets routed work, it gets reviewed, and it leaves a record. In XSI-AIMS, an open standard we author, that's literal: a team of agents is structured the way a team of people is, with ten defined roles, a human at the top of the authority chain, and an Orchestrator whose entire job is sending each task to the resource that fits it.
Meet us where we work
Coordination doesn't happen in a chat window bolted to the side of the company. It happens in Microsoft Teams and Slack, in Azure DevOps and Jira, in the wiki and the repo — and, more than anywhere, on calls. Those aren't incidental tools. They're the surfaces on which a team is a team. A human teammate meets the function where the function already operates. No one asks a designer to lay out a campaign in a spreadsheet. A coworker adapts to the team's tools — it doesn't make the team adapt to it.
So we started with the voice
One tool predates every other on that list. Teams talk. They interrupt, they redirect, they say "not like that" before the work goes too far in the wrong direction. A large part of how any team coordinates is never typed anywhere — it's spoken, in a meeting or on a call. An agent you can only reach by typing into a box is held permanently outside that channel. That's where we decided to start.
In preview now, an agent can join a meeting by speech and take a huddle — answer, and be interrupted, live, the way a colleague would. It runs on your own infrastructure. XSI LodeStone is a sovereign appliance built on kernel-isolated orchestration, so the conversation and the oversight of it stay inside your environment, not someone else's cloud. And it's governed by default — the agent on the call answers to the same authority chain, the same Witness Layer where every action leaves a trace, and the same deterministic shadow layer that runs underneath the agents and can't be touched by the agents it watches. What changed is the surface. The accountability didn't.
What we're not claiming
Gap: Voice is in preview, not general availability — an early surface with a defined scope, open to a limited audience. Native participation across Teams, Slack, Azure DevOps, Jira, the wiki, and the function-specific tools of marketing, sales, and finance is the destination, not the current state.
I won't pretend it is done. We built the governance before the reach, on purpose. The oversight has to be real before the participation is safe — that ordering is the whole bet, not a gap we're apologizing for.
The hire who finally showed up
Come back to the hire who wouldn't use the tools. Nothing about their ability had to change for them to become part of the team. What had to change was presence — standing up, walking over, joining the meeting, taking the call. Becoming, finally, someone the team works with instead of around.
That's the whole aim. Not an agent smarter than the people it works with — an agent present alongside them, on the same surfaces, under the same oversight, so the work comes out better than either would manage alone. Voice is where that presence begins, because it's the surface teams have used the longest.
It's in preview, by invitation and waitlist. If your agents do real work and you've never once heard them in the room — come see what that's like. The full argument is in the companion whitepaper, Meet Us Where We Work.
Q&A with Rhyan
Extended questions from the argument above.
The surface. The hard part was never the speech — it was making an agent that can be present on a live channel, answering and being interrupted in real time, while it stays inside the same authority chain, the same Witness Layer, and the same deterministic shadow governance as every other agent we run. A voice agent that escaped the governance would be worse than no voice agent at all. So the work was bringing the oversight onto the new surface intact, not bolting a microphone onto an ungoverned bot.
Not by design. It runs on your own infrastructure. XSI LodeStone is a sovereign appliance built on kernel-isolated orchestration, so the conversation and the oversight of it stay inside your environment rather than someone else's cloud. For the organizations that drove this — regulated, security-conscious, sovereignty-bound — that is the whole point. The agent that joins the call is the agent that lives inside your boundary.
Because voice is the oldest coordination surface a team has, and the one the most coordination still runs through. A large share of how a team aligns is never typed anywhere — it's said in a meeting, on a call, in the moment someone redirects the work before it goes wrong. Text channels matter, and native participation in Teams, Slack, Azure DevOps, and Jira is on the roadmap. But an agent reachable only by typing is held permanently outside the channel teams use most. We started where the gap was widest.
Honesty about the boundary. Preview means an early surface with a defined scope, offered to a limited audience by invitation and waitlist — not general availability, and not a finished claim. The broad, native tool participation is the destination the category demands, and it is not done yet. What is solid today is the governed substance underneath — the routed roles, the authority chain, the audit and the shadow layers. The surfaces that substance reaches are catching up to it, in that order, on purpose.
Common Questions About XSI's Voice Support
Short answers to the questions teams ask first.
Conversational voice support lets a governed agent join a meeting by speech and take a huddle — answer and be interrupted in real-time conversational flow, the way a colleague would. It is available in preview for XSI LodeStone and XSI-AIMS for the Microsoft Agent Framework, and runs on the customer's own infrastructure.
No — it is in preview. An early surface with a defined scope, offered to a limited audience by invitation and waitlist. Scope and availability are subject to change during preview.
XSI LodeStone is XSI's sovereign agentic appliance and operating environment for running governed agentic coworkers on a team's own infrastructure, built on kernel-isolated orchestration so the work and its oversight stay inside the customer's environment.
It is XSI's runtime governance for agents built on the Microsoft Agent Framework — bringing governed, voice-capable resources to teams already building on it. XSI-AIMS is XSI's open agent-governance standard.
By the same machinery as any other agent XSI runs — an authority chain with a human at the top, a Witness Layer under which every action leaves a trace, and a deterministic shadow-governance layer that runs underneath the agents and cannot be modified by the agentic layer it watches. Voice changes the surface, not the accountability.
Extended Systems Intelligence Corporation (XSI) is an AI research and product development company. XSI authors XSI-AIMS, an open standard for governing autonomous AI agents, and builds the XSI LodeStone sovereign agentic platform. It is based in McCall, Idaho.