agencyteam — An expert orchestration framework for OpenClaw

🌐 English | 简体中文 | 繁體中文 | 日本語 | 한국어

agencyteam turns specialist agent workspaces into a reproducible expert orchestration workflow for OpenClaw.

It helps you build practical multi-agent workflows for code review, security review, architecture feedback, product critique, prompt engineering, and other AI agent automation tasks without hand-wiring every specialist yourself.

agency-agents refers to the upstream public expert roster from msitarzewski/agency-agents, which agencyteam converts into OpenClaw-ready workspaces and installable specialists.

It is designed for:

• parallel specialist reviews
• builder/reviewer splits
• code + security + product synthesis
• on-demand expert installation from the agency-agents upstream roster
• repeatable agent routing and expert workflow automation for OpenClaw

What it does

• converts supported upstream expert prompts into OpenClaw workspaces under ~/.openclaw/agency-agents/<agent-id>/
• registers those experts in agents.list
• preserves non-agency agents already present in your config
• merges installed expert IDs into main.subagents.allowAgents (and preserves an existing ['*'] wildcard if you already use one)
• provides scripts for install, update, and on-demand spawn
• pins the default upstream source revision via UPSTREAM_REF for reproducible installs
• includes GitHub Actions smoke testing for script regressions

Built for teams that want faster multi-expert execution without giving up deterministic installs, safer config sync, or clean OpenClaw integration.

How it fits together

flowchart LR
  U[agency-agents upstream\npublic expert roster] -->|convert.sh| W[OpenClaw workspaces\n~/.openclaw/agency-agents/<agent-id>/]
  W -->|sync_openclaw_config.py| C[openclaw.json\nagents.list]
  C --> A[main.subagents.allowAgents]
  W --> S[install.sh / update.sh / spawn-and-install.sh]
  A --> R[On-demand expert runs]
  S --> R

Loading

• convert.sh transforms upstream expert prompts into a staged snapshot of local OpenClaw workspaces
• install.sh syncs that staged snapshot into the live workspace root, then registers selected experts and updates main.subagents.allowAgents
• update.sh refreshes generated experts from a staged snapshot while preserving non-agency agents in your config
• spawn-and-install.sh repairs a missing or unhealthy expert from a staged snapshot before launching the workflow

Common use cases

• run parallel code review, security review, and architecture review on the same repo
• split builder and reviewer roles across specialist agents for safer implementation loops
• install expert personas on demand instead of preloading a large agent roster manually
• create reusable AI workflow patterns for engineering, product, design, QA, and growth work
• orchestrate prompt engineering and system prompt review with distinct expert perspectives
• build OpenClaw automation flows that stay reproducible across installs, updates, and gateway restarts
• support solo builders who want an "AI team" feel without maintaining complex custom routing by hand

Requirements

• openclaw
• git
• python3

Quick start

git clone https://github.com/siubing05/agencyteam-openclaw.git \
  ~/.openclaw/workspace/skills/agencyteam

cd ~/.openclaw/workspace/skills/agencyteam
./scripts/install.sh

Install options

Install all supported experts:

./scripts/install.sh
# or
./scripts/install.sh --all

Install only specific experts:

./scripts/install.sh --agents "engineering-code-reviewer engineering-security-engineer design-ui-designer"

Pinned upstream policy

By default, conversion uses the commit recorded in UPSTREAM_REF.

• this makes installs reproducible
• this narrows supply-chain drift compared with tracking main implicitly
• you can override it deliberately with AGENCYTEAM_UPSTREAM_REF=<tag-or-commit> or ./scripts/convert.sh --ref <tag-or-commit>

What the installer actually does

1. clones the upstream agency-agents repo to a temporary directory
2. converts supported categories into OpenClaw workspaces
3. syncs matching entries into agents.list
4. merges installed IDs into main.subagents.allowAgents
5. creates a timestamped backup of openclaw.json
6. restarts the gateway and waits for it to respond

Update workflow

Preview upstream changes without writing anything:

./scripts/update.sh --dry-run

Apply upstream changes:

./scripts/update.sh

Also remove agencyteam-managed agents that disappeared upstream:

./scripts/update.sh --prune-removed

Important update behavior

• update.sh refreshes generated AGENTS.md files from upstream
• if you manually edited generated files under ~/.openclaw/agency-agents/, an update can overwrite those edits
• non-agency agents in your config are preserved
• --prune-removed is opt-in so removals are explicit
• prune flows only remove directories marked with AGENCYTEAM_MANAGED; unrelated local directories under the same root are left alone
• config sync fail-fast validates malformed main.subagents.allowAgents values instead of silently dropping bad entries

On-demand installation + spawn

If you want to use one expert immediately and install it if missing:

./scripts/spawn-and-install.sh engineering-code-reviewer "Review this repository for correctness and maintainability" --timeout 600

Verify

openclaw agents list
openclaw gateway status

Advanced environment overrides

AGENCY_DEST=/tmp/agency-agents ./scripts/install.sh --agents "engineering-code-reviewer"
OPENCLAW_CONFIG_PATH=/tmp/openclaw.json ./scripts/update.sh --dry-run
AGENCYTEAM_UPSTREAM_REF=<tag-or-commit> ./scripts/install.sh
./scripts/convert.sh --ref <tag-or-commit>

Suggested public positioning

Use agencyteam when you want OpenClaw to behave like a small expert panel instead of a single generalist assistant.

CI

GitHub Actions runs .github/workflows/ci.yml, which performs:

• bash syntax checks
• python syntax checks
• a deterministic local smoke test with a fake openclaw shim and local upstream git repo

See also:

• SKILL.md
• references/routing.md
• references/workflows.md