Cortx turns single agent runs into larger units of work. One agent hands a sub-task to another, a run fans out into parallel steps, a workflow strings agents and people together with maker and checker steps, and a supervisor drives a team of agents through a plan. Every hand-off moves work, never authority.
People, schedules and outside agents start the work. Cortx composes it six ways, under one set of controls.
Every composed run, whether a child, a delegate, a workflow step or a member of a supervisor's team, holds no more authority than its caller held and it was granted, passes the same per-tool controls as a top-level turn, and is owned by a worker that outlives whoever started it.
Four steps run around every composed run. Each step hands the next what is on its arrow. Select a step to jump to it.
Six ways to compose work: sub-agents, A2A hand-offs, agents as tools, parallel steps, workflow graphs and supervisor runs.
Two engines divide the work. A durable workflow engine carries work across agents, people and days; the agent runtime carries the inside of each run: plan, act, reflect, steer, park.
Once armed, each hand-off is decided against intersecting scopes, depth, cycles and allowlisted edges. Each tool call passes the guardrail chain. Each checker step is a human task with an expiring, single-use approval link.
Hand-offs land on the ledger, delegation chains are measured on every A2A call, and the run graph and workflow history each record their half.
A composition that fails any one of them is not governed, however good its parts are.
Every orchestration control ships in observe mode first, so you see what it would decide before you arm it.
Workflows put people in the loop. A checker step is a human task with an expiring, single-use approval link, and a quorum counts distinct, authenticated people, not names in a payload.
A delegate gets what was requested, intersected with what its caller held and what the callee allows. The callee re-decides from a signed lineage, and outside agents get no user token, history or state.
Teams build agents in five steps in the Agent Catalog, without code. Their agents load through the same loader and tool wrapper as the agents we ship.
How authority, control and ownership hold when agents hand work to each other.
A delegate's authority is computed from what was requested, what its caller held and what it was granted. A forwarded user token is a credential for one hop, never a grant.
A tool reached through a sub-agent, an A2A hand-off or a workflow step passes the same control chain as a tool called in chat.
An agent outside Cortx gets no user token, no history and no state, only an opaque session id. If token exchange fails, the call is refused, never downgraded.
Every orchestration control ships off and observable, with a counter you can read, so you see what it would have decided before you switch it on.
We'll walk through a maker-checker workflow and an A2A hand-off on a scenario that looks like your firm.
We'd like to use Google Analytics to see which pages are read, how visitors find us, and roughly where in the world they are. Nothing is stored on your device unless you turn it on, and you can change your choice at any time. See our Cookie Policy for details.
What we collect