Tue, Sep 01, 2026 • News
One agent handing work to another
Anthropic shipped dynamic workflows in Claude Code. The nearest thing in AGENTEXA is not coming — it is already in every installation, and it is bounded on purpose.
Anthropic has shipped dynamic workflows in Claude Code. In their description, Claude writes an orchestration script for the task at hand: it plans, splits the work into subtasks, hands them to subagents running in parallel, checks the result before merging, and returns one answer. Coordination lives outside the conversation thread, progress is checkpointed, and an interrupted run continues rather than restarting.
That is worth reading in full, and it is not what this note announces. It describes an assistant composing its own helpers for a single task. What people ask us about is a different thing that sounds the same: can one agent hand work to another inside AGENTEXA?
It already does, in every edition
Delegation is part of the base delivery. It is not a feature of the expensive edition, it is not an add-on, and there is nothing to wait for. An agent can be given work by a person, by a schedule, by an incoming message — or by another agent. The conversation that a delegated run creates is an ordinary conversation with its origin recorded, so the work is visible where every other piece of work is visible.
What bounds it
The interesting part is the limits, because they are what make delegation safe to leave switched on.
A subordinate gets a subset, never a union. The capabilities handed to a subordinate run are drawn from the parent’s own set and default to read-only. An agent cannot reach an irreversible action by delegating to something that has more rights than itself.
Approval happens at the step that reaches the irreversible action, not at the top of the tree. A delegated run has no human initiator, so the request for approval carries the automation and its owner instead, and goes to the administrative role of that organisation. No agent approves on a person’s behalf, at any depth.
Depth is declared and finite. Recursion is either forbidden or capped at a stated depth, and the cap is part of the published limits of an installation.
Losing the author stops the branch. If the person who owns an agent loses access, that agent’s automations pause and its delegated personal authorisations are revoked. The work stops at that link and says so, rather than continuing under rights nobody holds any more.
What we are not building
We are not shipping a canvas where you draw a process and connect agents to each other. Multi-step work here is expressed as agent instructions, automations and connector bindings — the same material an author already edits — and not as a user-drawn graph. That is a deliberate exclusion, not a gap waiting to be filled. If it ever changes, it will change as a scope decision with a written trail, and this page will say so.