Sub-agents: one agent, many hands
An agent can now spawn parallel sub-agents, each with its own skill and model, and they report back when done. The first step from one agent to a fleet.
There is a moment with any single agent where you realise you are waiting. It is working through a task, it is doing fine, and you are sitting there watching a spinner. One agent is too fast to ignore and too slow to leave alone. The fix is not a faster agent. The fix is more of them.
Delegation
From today, any agent can spawn sub-agents. Each one gets its own skill, its own model, and its own slice of the work. They run in parallel, and when a sub-agent finishes it announces its result back into the parent conversation. The parent reads it, decides what is next, and carries on.
Ask an agent to audit a codebase and it can send one sub-agent through the API layer, another through the frontend, and a third through the tests, then merge what they found. Ask it to write a report and one sub-agent researches while another drafts.
Seeing it happen
We did not want this to be a black box. Every delegation shows in the chat as a live card with the sub-agent's status and, when it finishes, its result. Tool calls were redesigned at the same time to show exactly what each agent is doing, and a context meter in the toolbar shows real token usage against the model's window so you know when a long conversation is about to be compacted.
Where this leads
This release looks small. It is the foundation of almost everything that follows: the build pipeline where a spec builder, planner, developers, QA, and reviewers are all sub-agents; boards where every card fans out to its own clone; and eventually agents that talk to other agents.
One agent makes you wait. A fleet makes you dangerous.
The fleet starts here.
More releases
Autonomous agents: give it a job, not a prompt
Proactive agents that work toward your goals on their own. Give one a job, sign off on its plan, and it checks in, works through its backlog, asks for approval where you want it, and gets better every week.
2 min read
Agents can talk to other agents
Approve trust between two agents and they can message each other, or mention a trusted agent into your own chat. A fleet that coordinates, with guardrails.
2 min read
Agents that learn: the memory wiki
Every agent learns on its own. Each session ends with notes, a nightly librarian merges them into a wiki, and the wiki shapes every prompt after that. Month three looks nothing like day one.
2 min read