Changelog2026-10-04New

Parallel agents with AgentGroups (Beta)

Parallel agents are in Beta. An AgentGroup runs many AgentRuns of one agent: you create the group, add a run for each task, and the group releases them as slots and budget allow.

• Create runs for a group. Give each run spec.group and a spec.taskKey. nodus get ag shows each group’s phase, finished runs out of all, active runs and cost, and nodus get ar --field-selector spec.group=NAME lists its runs. status.blockedReasons says why queued runs have not started.
• Limit how many run at once. spec.limits.maxActive (default 10) is how many runs go together, and spec.limits.maxPending (default 10000) is how many unfinished runs the group takes in.
• Order runs with dependencies. A run with spec.dependsOn starts after those tasks succeed. When one fails or is cancelled, the run is cancelled with reason DependencyFailed or DependencyCancelled.
• One cost cap for the group. With spec.maxCostUSD, a run starts only while its own cap still fits under what is left, so the runs together never spend more than the cap.
• Seal, cancel and delete. Sealing a group says that no more runs are coming, and a sealed group finishes once its runs have. nodus cancel ag/NAME cancels every unfinished run, and deleting the group deletes its runs.
• Batches. POST …/agentruns takes an AgentRunList of up to 100 runs, created together or not at all.
• From Python. nodus.AgentGroup.create(...) and group.submit_many(...) create a group and its runs, ordered by their dependencies, and agent.map(inputs) runs one task per input and yields the answers. These no longer raise nodus.errors.Unsupported; max_held=, evaluation= and group.results() still do.

Evals are not available yet. See Run agents in parallel.