SuperAPI AI Directory
Partner中文AI newsFavorites
Skillsagent-team-manager

agent-team-manager

Agents & models

View repository

Updated 10-15 16:44

About this skill

Multi-agent team management

Manage CountBot multi-agent teams via the command line, covering team CRUD, member CRUD, and team-level model configuration.

Use cases

  • User says "help me create a multi-agent team" -> create team
  • User says "make a document deep analysis team" -> create team by template
  • User says "change this team to dependency graph mode" -> modify team
  • User says "add a reviewer role to the team" -> add member
  • User says "change the analyzer's task" -> modify member
  • User says "optimize this role's prompt" -> first look at the current member configuration, then modify task and/or system_prompt
  • User says "delete the summarizer role" -> delete member
  • User says "configure a separate model for this team" -> configure team custom model
  • User says "show me what teams/members there are" -> list or detail

How to invoke

All operations are executed via the exec tool:

python3 skills/agent-team-manager/scripts/agent_team_manager.py <command> [args]

Strict syntax

  • Only these first-level commands exist:list、info、template-list、create、update、delete、member-list、member-add、member-update、member-delete、config、config-set、config-reset
  • info / member-list / update / config-set 's team parameters are all positional parameters; --team
  • member-update 's member parameter is the second positional parameter member_ref, which does not support --id
  • To toggle the skill system, use the team command update "团队名" --enable-skills or --disable-skills
  • First look at --help, then execute according to the positional parameter order in the help; do not invent subcommands or parameter names on your own

Common commands

List teams

python3 skills/agent-team-manager/scripts/agent_team_manager.py list

View team details

team_ref Supports team name, full ID, ID prefix.

python3 skills/agent-team-manager/scripts/agent_team_manager.py info "文档深度分析"

Create team

# 创建空团队，后续再逐个添加成员
python3 skills/agent-team-manager/scripts/agent_team_manager.py create \
  --name "文档深度分析" \
  --description "理解文档 → 提取要点 → 分析问题 → 生成总结报告" \
  --mode pipeline \
  --enable-skills

# 直接按内置模板创建
python3 skills/agent-team-manager/scripts/agent_team_manager.py create \
  --template document-analysis

Modify team

python3 skills/agent-team-manager/scripts/agent_team_manager.py update "文档深度分析" \
  --mode graph \
  --description "先并行抽取，再汇总结论" \
  --active

python3 skills/agent-team-manager/scripts/agent_team_manager.py update "文档深度分析" \
  --disable-skills

Delete team

python3 skills/agent-team-manager/scripts/agent_team_manager.py delete "文档深度分析"

Member management

List members

python3 skills/agent-team-manager/scripts/agent_team_manager.py member-list "文档深度分析"

Add member

python3 skills/agent-team-manager/scripts/agent_team_manager.py member-add "文档深度分析" \
  --id reader \
  --role "文档理解专家" \
  --task "通读文档，理解整体结构和核心内容" \
  --system-prompt "你是文档理解专家，先识别文档结构，再提炼核心主题。"

Graph mode can include dependencies and conditions:

python3 skills/agent-team-manager/scripts/agent_team_manager.py member-add "代码检查团队" \
  --id refactor \
  --role "重构建议专家" \
  --task "基于前序检查结果给出重构建议" \
  --depends-on syntax-checker,logic-analyzer \
  --condition-type output_contains \
  --condition-node logic-analyzer \
  --condition-text 严重

Council mode recommends filling in --perspective：

python3 skills/agent-team-manager/scripts/agent_team_manager.py member-add "投资评审会" \
  --id risk \
  --role "风险分析师" \
  --perspective "风险与合规"

Modify member

python3 skills/agent-team-manager/scripts/agent_team_manager.py member-update "文档深度分析" analyzer \
  --task "重点分析论证链、信息缺口和潜在偏见" \
  --system-prompt "你是批判性分析专家，输出问题、证据和风险。"

Modify an existing role prompt (strong constraint)

Whenever the user expresses the following intentions, handle them uniformly as "modify an existing role prompt":

  • Optimize role prompt
  • Change prompt / change system prompt
  • Adjust role setting, tone, boundaries, output requirements
  • Make a member "more professional / stricter / more like a certain type of expert"

The mandatory execution order is as follows.

1. First read the current real configuration; modifying based on impressions is not allowed

python3 skills/agent-team-manager/scripts/agent_team_manager.py info "文档深度分析"

You must first check whether the target member currently has task、system_prompt, then decide which field to modify. It is forbidden to directly generate member-update commands before reading the current configuration.

2. Field attribution determination rules

  1. If the target member only has taskand no system_prompt
  2. it means the team stores the main prompt body directly in task
  3. At this point, when the user says "change the prompt", by default prioritize modifying --task
  4. If the target member already has system_prompt
  5. when it involves role personality, expert identity, tone, principles, boundaries, or long-term behavior constraints, prioritize modifying --system-prompt
  6. When it involves specific work content, execution steps, output structure, delivery format, or check items, prioritize modifying --task
  7. If the current task is itself an entire prompt-style text, and the user wants a systematic refactor, you should also modify --task and --system-prompt

3. Execution Specifications

  • Modifications must be persisted; the final action must be to execute member-update
  • member-update The correct form is: python3 skills/agent-team-manager/scripts/agent_team_manager.py member-update "团队名" 成员ID [flags]
  • Member IDs use positional parameter member_ref, do not write it as --id
  • After the modification is complete, it is recommended to execute info "团队名" again to review the result
  • If the user requests to "modify the prompt", but the current configuration only has task, do not only modify system_prompt
  • If the user requests "make the system prompt more professional", but the old long prompt still remains in task , you should determine whether task

Example: the current member only has task, without system_prompt

python3 skills/agent-team-manager/scripts/agent_team_manager.py member-update "小红书文案团队" reviewer \
  --task "审核小红书文案，检查敏感词、风格统一、内容完整、平台规范和可读性；输出审核结论、修改建议、最终发布版与发布提醒。"

Example: explicitly modify the system prompt

python3 skills/agent-team-manager/scripts/agent_team_manager.py member-update "小红书文案团队" reviewer \
  --system-prompt "你是严格的小红书内容审核专家，优先识别违规风险和夸大表达，输出结论必须清晰、克制、可执行。"

Example: split the old prompt-style task into "task + system prompt"

python3 skills/agent-team-manager/scripts/agent_team_manager.py member-update "小红书文案团队" reviewer \
  --task "审核小红书文案并输出审核结论、修改建议、最终发布版与发布提醒。" \
  --system-prompt "你是严格的小红书内容审核专家，重点检查敏感词、极限词、医疗宣称、风格统一、内容完整和平台规范。"

4. Professional Handling Principles

  • The goal is to "modify the effective configuration", not to "output a piece of copy that looks better"
  • First identify the existing data structure, then decide which field to modify
  • Follow the principle of minimal necessary modification, avoiding merely adding fields while retaining old conflicting content
  • If the user does not specify a field name, the word "prompt" should be interpreted in light of the current storage structure and must not be mechanically equated with system_prompt

Delete member

python3 skills/agent-team-manager/scripts/agent_team_manager.py member-delete "文档深度分析" summarizer

Team model configuration

View current configuration

python3 skills/agent-team-manager/scripts/agent_team_manager.py config "文档深度分析"

Set custom model

python3 skills/agent-team-manager/scripts/agent_team_manager.py config-set "文档深度分析" \
  --provider zhipu \
  --model glm-5 \
  --temperature 0.5 \
  --max-tokens 8192

Reset to global default

python3 skills/agent-team-manager/scripts/agent_team_manager.py config-reset "文档深度分析"

Parameter description

  • mode only supports pipeline、graph、council
  • --enable-skills When enabled, sub-agents can read and use skills/*
  • --cross-review / --no-cross-review is only meaningful for council mode
  • depends_on is only meaningful for graph mode
  • perspective is mainly used for council mode
  • task will be passed to the sub-agent as the execution-phase # Your Task , and is the actual task description
  • system_prompt is the role's long-term persona/responsibility setting
  • If system_prompt is empty, the system will base it on role + task to automatically generate a default system prompt
  • If system_prompt has a value, it will be used directly as the system message; but task will still continue to be passed into the execution prompt
  • Therefore: when an existing member only has task , prioritize updating task; when a stable role tone/boundaries are needed, then add or modify system_prompt
  • Many historical teams wrote an entire "prompt-style description" directly into task, which is not a script failure, but rather how the data was originally stored
  • So when "the prompt was changed but did not take effect", first check whether the old prompt is still lying in task inside

Built-in template

Currently built-in:

  • document-analysis: in-depth document analysis,pipeline mode, skill system enabled by default

Notes

  • Team name must be unique
  • Member IDs must be unique within the same team
  • Modifying members essentially reads the team details and then writes back the whole agents
  • If the team has enabled a dedicated model, executing workflow_run(team_name="团队名", goal="...") will automatically inherit that model configuration
  • If the CountBot backend is not started, the script will directly report a connection failure