It is 7:40 on a Tuesday morning and Priya has a question her director wants answered by lunch: does a four-day work week really keep output up? She could paste it into a chatbot and get a confident paragraph back. But the same model that finds the numbers would also be the one deciding whether those numbers hold up, and it rarely disagrees with itself. What she wants is one person to find the trials, another to pull the figures, and someone else to send back anything without a source.
Two entrances to the same screen
osFoundry is a desktop app for running AI work on your own machine, and it has two closely related ways to hand work to agents. Agent (shown in the sidebar as Single agent) is one named agent with one purpose: you give it instructions, it gets its own ID such as AG-0001L, and every run it does is recorded so you can follow it later. Agent team puts a leader in front of a roster of specialist members. The leader plans the work, gives each step to a member, and a reviewer checks each step before it reaches you. Both live in the sidebar under Agents, right next to each other, and they open the same screen. The difference is the size of the roster. If you have used a group chat with a project manager in it, Agent team will feel familiar. The difference is that every message, handoff and verdict is kept, and you can scroll back through it.
A look around the team screen
This is a finished team session. The question and answer are example data from a demo run.
- The leader's conversation (centre). The header reads "Leader · Orchestrate your team". This is where you ask the question once and where the combined answer lands, here as a table with a source on every row, a "Where sources disagree" line and a "Not covered" line.
- Session tabs (top of the centre). Chat, Team mail, Email, Handoffs, Sub-agents, Proposals and Files. Handoffs shows who passed work to whom. Proposals holds changes the agents suggest to files or settings. A suggestion is not applied until it is accepted.
- Agent session panel (right). There is one card per member, with what the leader asked it to do, the tools it called (
web_fetch,web_search) and its result. Under each result you see the reviewer's verdict, such as "Approved". A run selector ("Run 1 · 07:40 AM") and a Cards / Graph switch sit on top. - Team roster tab. This is the list of who is allowed to act in this session. A team is a roster, not a workflow. The order of work comes from the leader's plan.
- The session rail (left, hidden here) and the inspector. The rail lists every session with its ID. The inspector has Goal, Trigger and Hooks tabs, for a finish line, what started the run, and the checks around the work.
Three ways people use it
A researcher who wants every number sourced
Priya opens Agent team and presses New session. In the New team session dialog she keeps "Start from a team" and picks Deep Research Team from the Team list. The dialog loads its five members (deep-research-agent, research-analyst, data-researcher, search-specialist, academic-researcher). Under "Who leads" it shows Orchestrator, which "plans the work up front, then delegates the steps to the agents below". She could press "make leader" on any member to put that member in charge instead.
She presses Create, types her question into the leader's chat and sends it. The leader splits it into five parts: find the trials, the studies, the numbers, the counter-evidence, then one answer. Three members start at once. The Data Researcher waits for the Search Specialist's list before it pulls figures. When the Data Researcher reports a 40% figure, the reviewer marks it "Sent back" with the reason that it has no source, date or measure. The step runs again, finds the source, and is approved.
What she ends up with is one answer in the centre pane, and every figure in it has passed a check by someone other than the agent that found it. She also gets an honest "Not covered" line she can paste into her note to the director.
A solo developer who keeps pasting the same instructions
Tom turns meeting notes into action items several times a week, and each time he pastes the same paragraph of rules into a chat first. That job does not need a team. It needs one agent that remembers the rules.
He opens Single agent, presses New session and stays on the Create in chat tab. In the "New agent" box he writes one line about what he needs and presses Start. The chat that opens is a setup agent. It asks a couple of questions, then proposes a name, an @handle and full instructions. A card appears above the composer with the name, the handle, a Show instructions link and a Create agent button.
He opens the instructions and checks that the rule he cares about most is in there: never guess an owner or a deadline. Then he presses Create agent. The same chat is now that agent. When he pastes his notes, missing owners and dates are marked as missing, never guessed. The session keeps its Agent ID in the rail, and the next time he only has to pick the agent.
A release lead who wants to see who did what
Dari runs releases for a small product team. Her team session has a leader and members named Release Drafter, Change Auditor, Docs Reviewer, QA Summarizer and Approval Prep. Each one owns a slice of the release package.
Before the run she opens the Goal tab and presses Set a goal. The panel explains that a goal is broken into success criteria that start failing and turn green as the agents satisfy them. The session keeps running until they all pass or it hits its round limit. She writes three criteria, the last one being that the approval packet is complete. While the members work, she checks Handoffs to follow the chain from the leader to each member and back. She also checks Proposals for any config change a member suggests, and nothing there is applied without her. She can steer the leader at any point by replying in its chat, and Pause in the panel header holds the run while she reviews.
At the end she has a release package plus a record of who drafted, audited and approved each part. She can show it to anyone who asks how a line ended up in the notes.
Where it leads in the rest of the app
Agent and Agent team are where work starts. Sessions is where you watch it. That board shows every chat, agent, team and terminal session as one card, sorted into Running, Needs you, Paused, Idle, Done and Failed by what the session is actually doing.
When the same team job has to run again and again, Loop wraps a roster with a trigger, limits and a round-by-round record of which criteria passed. Graphs are for work whose path depends on what happens along the way. The leader's conversation is a Maestro chat, so a team's answer can carry on into the rest of your Maestro work. The session's Enable sync link connects it to a project message channel, where people and agents can talk in one place. Read more on the Agents feature page and the agent sessions overview.
Questions newcomers ask
When should I use a single agent instead of a team?
Use a single agent when the job is one clear task with fixed rules, like reformatting notes or drafting replies in one voice. Use a team when the job has parts that need different tools, or when you want a separate reviewer to check the work before you see it.
Can I change the team after the session starts?
No. The New team session dialog says the team "is fixed for this session's life". Within the dialog you can still add or remove agents and choose who leads before you press Create. To use a different roster, start a new session.
Does any of this run in the cloud?
The New session form offers Local and Isolated runs. It notes that both run on your machine with your own keys and that nothing goes to the cloud. Chat history is local by default.
Can the agents rewrite their own instructions?
They can suggest changes, which show up under Proposals. A proposal is not accepted configuration. Whether agents may touch their own prompts is a separate setting.
You can get the desktop app at osfoundry.io/download.
