@agentskit/runtime

Profile: major-package

Run autonomous agents in 5 lines — no UI, no boilerplate, just results.

Tags: ai · agents · llm · agentskit · openai · anthropic · ai-agents · autonomous-agents · react-loop · orchestration · multi-agent

Verified proof

• Package metadata and tests live under packages/runtime/.
• Package guide: https://www.agentskit.io/docs/reference/packages/runtime
• Stability map: docs/STABILITY.md

How this fits the ecosystem

@agentskit/runtime is the headless agent engine: ReAct loops, tools, skills, delegation, memory, cancellation, and structured run results.

• AgentsKit: compose it with the other packages in this repo to build agents from small, swappable parts.
• Registry: look for ready agents and templates that already use this layer at registry.agentskit.io.
• Playbook: learn the production patterns behind this layer at playbook.agentskit.io.

Docs: package guide · agent handoff

Why runtime

• ReAct loop handled for you — observe, think, act, repeat: the runtime drives the full cycle and stops when the agent decides it's done
• Structured, inspectable results — every run returns the final content, step count, all tool calls made, and total duration; no black boxes
• Production-ready lifecycle — lazy tool init/dispose, AbortSignal cancellation, memory persistence, and AgentEvent emissions for observability
• Multi-agent delegation — pass a planner skill with named delegates and the runtime coordinates sub-agents automatically

Install

npm install @agentskit/runtime @agentskit/adapters

Quick example

import { createRuntime } from '@agentskit/runtime'
import { openai } from '@agentskit/adapters'
import { webSearch, filesystem } from '@agentskit/tools'

const runtime = createRuntime({
  adapter: openai({ apiKey: process.env.OPENAI_API_KEY, model: 'gpt-4o' }),
  tools: [webSearch(), ...filesystem({ basePath: './workspace' })],
  systemPrompt: 'You are a helpful research assistant.',
})

const result = await runtime.run('Research the latest advances in quantum computing')
console.log(result.content)
console.log(`Completed in ${result.steps} steps, ${result.durationMs}ms`)

With skills

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

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

const result = await runtime.run('Summarize REST vs GraphQL', { skill: researcher })
console.log(result.content)

Features

• createRuntime — single entry point for headless agent execution.
• ReAct loop: observe → think → act → repeat until done.
• Returns { content, steps, toolCalls, durationMs } — fully inspectable.
• AbortSignal cancellation support.
• Tool init / dispose lifecycle hooks.
• AgentEvent emissions for observability integrations.
• retriever option for RAG context injection.

Advanced primitives

Topology cancellation

AgentHandle.run(task, context?, signal?) receives cancellation separately from caller context; existing two-argument handles remain compatible. swarm has no per-member deadline unless timeoutMs is set, which must be finite, greater than 0, and no greater than 2,147,483,647 milliseconds. Invalid values throw ConfigError (AK_CONFIG_INVALID) synchronously. On deadline, the child signal is aborted and an optional legacy abort hook is called once on a best-effort basis; children must honor the signal to stop their work. If all members fail, swarm throws RuntimeError (AK_RUNTIME_DELEGATE_FAILED) with the first failure as its cause, including NetError (AK_NET_TIMEOUT) for a deadline. A child that ignores cancellation may continue after the swarm returns.

Ecosystem

Contributors

License

MIT — see LICENSE.

Docs

Full documentation · GitHub

Maturity and compatibility

• Stability: beta — see docs/STABILITY.md
• Node.js 20+ and TypeScript strict mode
• Published as @agentskit/runtime

Contributing

See CONTRIBUTING.md and the monorepo LICENSE.