@agentskit/runtime — for agents

#Purpose

Run an agent without a UI. Supports reflection, planning, multi-agent orchestration, durable step logs, cron + webhooks, speculative execution.

#Install

npm install @agentskit/runtime

#Primary exports

• createRuntime({ adapter, tools, memory, skills, observers, ... }) — one-line agent; runtime.run(task) returns { content, steps, ... }.
• createSharedContext — typed shared context across tools.
• speculate({ candidates, pick, timeoutMs }) — race adapters, abort losers. See Speculative execution.
• supervisor, swarm, hierarchical, blackboard — multi-agent topologies. See Topologies.
• createCompareHandler, createVoteHandler, createDebateHandler, createAuctionHandler — cooperative fan-out-then-select topologies. Each is generic over the run context, takes a plain config + an injected TopologyRunAgent, and returns a TopologyOutcome (ok / failed / paused). Compare selects by all/first/eval/judge/manual; vote tallies majority/weighted/unanimous/quorum ballots with a tie-break; debate runs a proponent/opponent/judge loop; auction picks the best bid by lowest-cost/highest-confidence/fastest/custom under a reserve price. Shared helpers: settleWithConcurrency, resolveConcurrency, DEFAULT_TOPOLOGY_CONCURRENCY, InMemoryScratchpadStore.
• createDurableRunner + createInMemoryStepLog / createFileStepLog — Temporal-style step-log durability. See Durable execution.
• createCronScheduler (5-field cron + every:<ms>) + createWebhookHandler + parseSchedule + cronMatches — background agents. See Background agents.
• compileFlow({ definition, registry }) + validateFlow + flowToMermaid — compile a YAML / object FlowDefinition into a durable DAG runner. See Visual flows.
• createChatTrigger({ adapter, agent, ... }) — unified inbound trigger for chat-surface bots (Slack / Teams / Discord / WhatsApp). Wraps a ChatSurfaceAdapter that normalizes provider events into a ChatSurfaceEvent discriminated union (message / mention / reply / reaction / file_upload / installation). Returns a framework-agnostic WebhookHandler.
• createQuotaTracker + withQuotas — per-tool quota / blast-radius limits (count / cost / duration windows).
• createValidatorGuard + built-in validators denyPattern, lengthRange, isJson — agent-insurance primitive that rejects tool args / outputs against allow/deny rules before they propagate.
• piiDenyValidator(opts?) — a Validator that fails when the output still contains PII (bridges @agentskit/core/security's redactor into the guard chain); a deterministic last-line gate for redaction/export agents instead of trusting the prompt.
• invokeStructured({ adapter, tool, task, parse, skill? }) — first-class structured output: run a skill that must call one submit_* tool, read it back from result.toolCalls, return the validated value. Replaces the hand-rolled "offer one tool, find the call, parse args" boilerplate every pipeline agent writes.

#Minimal example

import { createRuntime } from '@agentskit/runtime'
import { anthropic } from '@agentskit/adapters'

const runtime = createRuntime({
  adapter: anthropic({ apiKey: process.env.ANTHROPIC_API_KEY!, model: 'claude-sonnet-4-6' }),
})

const result = await runtime.run('Summarize the quarterly report.')
console.log(result.content)

#Common patterns

• Survive crashes: wrap side effects in runner.step(id, fn) (durable).
• A/B across models: speculate or replayAgainst.
• Compose agents: supervisor / swarm / hierarchical / blackboard.
• React to events: createWebhookHandler + createCronScheduler.

#Related packages

• @agentskit/core
• @agentskit/adapters
• @agentskit/skills
• @agentskit/observability

#Source

• npm: https://www.npmjs.com/package/@agentskit/runtime
• repo: https://github.com/AgentsKit-io/agentskit/tree/main/packages/runtime