Skip to content

About

Experimental TypeScript agent core with composable plugins.

Resources

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

agent-core

English | 简体中文

QuiFox: a Harness built on agent-core

If you are interested in agent-core, start with QuiFox to see how the framework and its plugins come together in an interactive Agent application.

QuiFox is a Harness built on agent-core, with terminal (TUI) and desktop (GUI) interfaces. It brings model connections, conversations, tool execution, approvals, Skills, child Agents and execution traces into one interactive environment.

Read the QuiFox guide →

The workflow workbench runs reusable TypeScript multi-Agent scripts, with static call graphs and live invocation states.

Agent Core

agent-core is an Agent framework built around a minimal execution core and explicit plugin composition.

Core provides a default execution loop and Hooks across its lifecycle. Plugins supply model generation, sessions, prompts, Tools, interruption, recovery, compaction, events and collaboration, so developers can compose their own Agent behavior.

Agent Core is currently experimental. APIs and plugin composition contracts are evolving.

Design demo

A visual tour of lifecycle Hooks, sessions, interruption and recovery, plugin composition, tracing, Code Mode, and peer-Agent collaboration, followed by application examples.

Watch the Agent Core and plugin capabilities demo.

Explore plugin composition

  • Keep conversation history with Chat or Responses sessions.
  • Combine interruption, checkpoints, and new input during execution.
  • Share a Teams plugin among ordinary peer Agents.
  • Execute Host-provided code and recompose plugin sets.

Try from source

With the Bun version pinned in the root manifest, install and build the workspace from the repository root:

bun install --frozen-lockfile
bun run build

DeepSeek Chat example

A Bun Host can compose Chat history and model generation as one conversation plugin. This example uses the DeepSeek deepseek-flash model through Chat Completions (official model documentation).

This is one example of plugin composition. You can implement and combine plugins to create richer capabilities and different agent behaviors.

Run the following JavaScript from the repository root. The Host supplies DEEPSEEK_API_KEY through its environment.

import { createAgent } from '@quifox/agent-core'
import { ChatModelPlugin } from '@quifox/agent-plugin-model'
import { ChatSessionPlugin, ChatConversationPlugin } from '@quifox/agent-plugin-session'

const session = new ChatSessionPlugin({ id: 'example', ownerId: 'example' })
const generation = new ChatModelPlugin({
    session,
    model: 'deepseek-flash',
    baseUrl: 'https://api.deepseek.com',
    apiKey: process.env.DEEPSEEK_API_KEY,
    assistantReasoningField: 'reasoning_content'
})
const conversation = new ChatConversationPlugin({ session, generation })
const agent = createAgent({ plugins: [conversation] })
const outcome = await agent.run({ prompt: 'Reply with exactly "Hello World".' })
if (!outcome.ok) { throw outcome.error }
console.log(conversation.readOutput())

Packages

Package Responsibility
agent-core Default loop, Hooks, Context, Outcome, and Tool contracts
agent-plugin-model Model generation Hooks and internal protocol/transport adapters
agent-plugin-session Chat history, storage, fork, and Conversation composition
agent-plugin-compaction Context budgets, summaries, and checkpoints
agent-plugin-checkpoint Current-turn execution records and recovery data
agent-plugin-interrupt Cooperative pause and terminate policies
agent-plugin-steering Input batching and run handoff
agent-plugin-code-mode Host code execution and asynchronous Tool proxies
agent-plugin-composition Plugin-array proposals and next-Agent reference handoff
agent-plugin-tools Tool catalog, visibility, and entry routing
agent-plugin-system-prompt Caller-provided trusted prompt contributions
agent-plugin-subagents Child Agents, fork, messages, waiting, and stop
agent-plugin-tasks Background operations and factual task receipts
agent-plugin-project-context Standard Skills and scoped AGENTS.md
agent-plugin-request-context Temporary model context separate from original history
agent-plugin-memory Two-stage memory and source verification
agent-plugin-teams Peer mailboxes and member wake-up scheduling
agent-plugin-events Per-Agent execution observations
agent-plugin-formal-backend Reference Responses backend composition

Examples and further reading

Related projects

agent-core assembles capabilities through lifecycle Hooks around a minimal execution kernel. DeepSeek Harness (DSH) likewise explores composable, plugin-based agent systems, with its plugin architecture built on Cordis.

About

Experimental TypeScript agent core with composable plugins.

Resources

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages