Skip to content

Multi-agent orchestration

multi-agent coordinates parent runs and child subtasks through an injected AgentRunner (the AgentEngine). Child runs use the same Action, Observation, policy, memory, and execution-log contracts as any single-agent run.

  • SubagentDelegationManager implements the core DelegationManager contract: spawn_subagent runs a child immediately or as an async subtask, join_subagent runs/observes it, cancel_subagent stops it. Children inherit parent tools and capabilities, get bounded budgets, and return their result as a parent observation.
  • TaskGraphCoordinator adds append-only handoffs, child resume/cancel hooks, observer trace bridging, budget summaries, and shared-memory write proposals + authorization.

Orchestration patterns

Reusable patterns composed purely over the AgentRunner seam (no kernel change). Each takes a runner and OrchestrationSteps ({ agent, task, host, budget? }):

Pattern What it does
runParallel(runner, steps, { concurrency }) Bounded-concurrency fan-out, results in input order; a throwing branch becomes a failed result instead of sinking the batch.
mapReduce(runner, steps, reduce, { concurrency }) Fan out, then combine the results.
vote(runner, steps, { answerKey, concurrency }) Debate/plurality over N runs (same task with different personas, or N samples) β†’ VoteResult { result, key, agreement, results }.
plannerWorkerCritic(runner, { planner, decompose, synthesize }) One planner run decomposes the task, workers run sub-steps in parallel, a critic synthesizes. The host supplies decompose/synthesize. 
import { vote } from "@agentic-kernel/multi-agent";

const result = await vote(engine, [stepA, stepB, stepC]);
console.log(result.key, result.agreement);

from agentic_kernel.multi_agent import vote

result = await vote(engine, [step_a, step_b, step_c])
print(result.key, result.agreement)

These patterns are validated under real models in the comprehensive real-LLM campaign.