A lightweight framework to empower agentic execution of complex tasks.
Compose models, tools, skills, and subagents. Customize the harness for context, permissions, planning, execution, and verification to build your own productivity agent.
01 / COMPOSABLE AGENTS
Configure roles, models, and tools for specialized agents. Combine task delegation with workflow orchestration to coordinate complex work.
Coordinate exploration, development, and research specialists while managing overall progress and results.
# general/agent.yaml
subagents: [explore, build, research]
callbacks:
- state_inject
- loop_guard
- subagent_limit
subagent_limit:
max_parallel: 4Review requirements and project context, then check that the plan covers the task.
# plan/agent.yaml
callbacks:
- state_inject
- plan_check
- loop_guard
tools:
todo_list:
mcp: falseCross-check sources and develop grounded conclusions, with checks for report completeness.
# research/agent.yaml
callbacks:
- state_inject
- round_reminder
- stop_gate
stop_gate:
enabled: true
max_retries: 2Use read-only tools to locate and analyze information for planning and execution.
# explore/agent.yaml
tools:
file_system:
mcp: false
include:
- read_file
- grep
- globWrite, run, and diagnose code, then use feedback to refine and verify the result.
# build/agent.yaml
tools:
file_system:
mcp: false
include: [read_file, write_file,
edit_file, grep, glob]
code_executor:
mcp: false
implementation: python_env02 / COMPOSABLE HARNESS
Combine context management, permissions, and quality checks to shape agent execution. Use lifecycle hooks to bring your own rules into planning, collaboration, and feedback.
Organize multi-step work with todos and check that the plan is complete and feasible before acting.
callbacks:
- plan_check
- state_injectCall tools or delegate focused tasks. Tool boundaries and concurrency limits shape each execution step.
callbacks:
- subagent_limit
- loop_guardInspect artifacts, pending todos, and answer quality. Return specific feedback so the agent can address what is missing.
callbacks:
- todo_gate
- stop_gateFinish after checks pass. Round reminders, retry limits, and budgets help bring extended tasks to a close.
round_reminder:
remind_before_max_round: 2
stop_gate:
max_retries: 2Plan · Act · Check · Finish
03 / LONG-HORIZON EXECUTION
Coordinate ongoing work with adaptive scheduling, layered permissions, and project memory, reducing manual supervision across multi-step and recurring tasks.
Run at fixed intervals or let agents schedule their next wake-up. Manage monitoring and maintenance tasks with limits on rounds, duration, and usage.
Loop / Cron · Fixed intervals and adaptive scheduling
Combine permission rules, model assessment, and human review. Approve routine requests automatically and route uncertain cases to a person.
Rules first · Per-call authorization · Human fallback
Organize history and compact long outputs
Keep project knowledge and preferences
Save progress and reconnect to the work
04 / BUILT-IN WEBUI
Manage projects, conversations, and files in one place. Follow plans and tool calls as your agent works, and review the results as they arrive.
MS-Agent / Agent WorkshopWEBUIWebUI recording · Qwen3.8-Max · Event materials to briefing and agenda
05 / APPLICATIONS
Specialized agents and workflows for research, software development, and video production, ready to adapt to your domain.
Combine search, evidence management, and report generation for open-ended research and document analysis.
├ src/
├ components/
└ tests/
Develop projects through requirements, architecture, implementation, and verification workflows.
Combine financial data, analytical methods, and reporting standards in specialized research workflows.
Orchestrate scripts, images, voice, and video tools for content generation and post-production.
06 / SKILL EVOLUTION
Analyze execution traces, evaluate outcomes, and revise skills through reflection. Validate candidates against tasks to retain useful improvements.
Explore projectIdentify execution issues
Generate a candidate skill
Retain effective changes
07 / OPEN ECOSYSTEM
Use your preferred models and community plugins. Connect tools and agents through standard protocols, and carry skills, instructions, and memory across frameworks.
Connect external tools or expose MS-Agent tools and applications as MCP services.
Connect ACP-compatible editors, call external agents, and route sessions through an agent proxy.
Call remote agents or publish MS-Agent as a discoverable service for other agents.
Load skills, commands, agent definitions, hooks, and MCP configuration from Claude Code, Codex, Cursor, and OpenClaw plugin formats.
Convert, merge, and sync instructions, skills, and memory. Share agent resources through ModelScope Hub and reuse them across frameworks.
08 / GET STARTED
Start in the WebUI, run tasks from the CLI, or integrate agents into your application with the Python SDK.
Read the docspip install -U "ms-agent[webui]"
ms-agent uiLaunch the workspace and configure your model.
pip install -U ms-agent
ms-agent run --query "Plan a research brief"Configure a model, then run a task from the terminal.
import asyncio
from ms_agent import LLMAgent
from ms_agent.config import Config
agent = LLMAgent(Config.from_task("agent.yaml"))
messages = asyncio.run(agent.run("Write a research brief"))
print(messages[-1].content)Define your agent in configuration and extend it with the SDK.
Full SDK example