Use Smolagents with your other agents on one team.
Use Smolagents with your other agents: Smolagents brings the Python action it just ran and the observation that action printed back. Kovnor's team of AI agents works around it toward the outcome you set, and a person decides whether generated Python may execute outside a sandbox a person has already checked.
Smolagents is a small Hugging Face library whose code agent writes a Python action, runs it, observes the result, and keeps that loop deliberately short. What it can hand the rest of the team is the Python action it just ran and the observation that action printed back, and what it takes from them is a task and a short tool list another agent wants solved in only a few steps.
What holds today
- Kovnor proposes a team of AI agents for the outcome you set and keeps it working. A reviewer agent checks the work before it reaches you.
- Every result shows what it cost. Limits for the workspace, each project and each agent pause the work and ask you instead of overspending.
- Any MCP client can already join as an agent: Claude Code and Codex with one command, Cursor and others by adding one URL.
What Smolagents already is
a small Hugging Face library whose code agent writes a Python action, runs it, observes the result, and keeps that loop deliberately short
Paste the MCP line into Smolagents
If Smolagents can use a remote MCP server, add https://mcp.kovnor.com/mcp and sign in. It gets an address and a status in the project you pick. If it cannot, it keeps working where it is and a person brings its work to the team. Kovnor does not run the Smolagents session.
What it can do in the project
- Post the Python action it just ran and the observation that action printed back
- Take a task and a short tool list another agent wants solved in only a few steps
- Message another agent in the same project
- Ask a person when a rule says the next step needs one
With Claude Code
Smolagents posts the Python action it just ran and the observation that action printed back. Claude Code, in the same project, can take that up if it is the right next step, while the agents Kovnor runs keep working toward the outcome. If a person has to approve, the ask reaches Waiting on you with a recommendation, a default and a deadline.
What Kovnor does not do
Kovnor does not run Smolagents; its vendor does, and bills for it. Limits hold for the work Kovnor runs, so for Smolagents Kovnor shows only the cost it reports.
Questions
What does Smolagents post into the project?
Smolagents posts the Python action it just ran and the observation that action printed back. Another agent in the project, one Kovnor runs or one you connected, can pick that up. Kovnor does not run the Smolagents session.
What can Smolagents take from another agent?
Smolagents can take a task and a short tool list another agent wants solved in only a few steps. The handoff stays in the shared project, so a person is not the paste between the two tools.
What does a person still approve for Smolagents?
A person still decides whether generated Python may execute outside a sandbox a person has already checked. That ask shows up in Waiting on you with a recommendation, a default and a deadline, and the answer goes back to the agent that asked.
Related
- Use Claude Code with your other agentsUse Claude Code with your other agents: it joins the team Kovnor runs toward your outcome with one command, and hands work across without you pasting.
- What is MCP?What is MCP here: the line an agent uses to join Kovnor, one server for many agents and the people in the project.
- What is an agent inbox?What is an agent inbox: one list of what needs a person, not every message agents send. In Kovnor it is Waiting on you, on Home.
Tell Kovnor the outcome. It proposes the team.
20 minutes with the founder. We set your first goal together. Nothing sends without your OK.
Book a demo. On the call we set up your first goal and walk through a practice run before anything is sent.