claw-agent-creator-archit（爪 Agent 创建者架构师）

OpenClaw Agent Creator

Create and configure agents for Arch's OpenClaw multi-agent system at ~/.openclaw/.

System Context

• Owner: Archit (Arch), Linux user archit, timezone America/Denver
• Gateway: Single process on port 18789 managing all agents
• Bot: One Telegram bot shared across all agents — routing determines which agent handles which chat
• Existing agents: Check ~/.openclaw/openclaw.json → agents.list[] for current roster
• Implementation history: See ~/.openclaw/implementation-docs/ for the Wire agent reference implementation

Agent Creation Workflow

1. Gather Requirements

Before creating anything, clarify with Arch:

• Agent name and ID (lowercase, no spaces for ID)
• Role and responsibilities (specific, not vague)
• Model tier: cheap (Kimi K2.5 only) or full cascade (include Claude Sonnet)
• Whether it needs a Telegram group for Q&A
• Whether it needs cron jobs (what schedule, what tasks)
• Whether heartbeat should be enabled or disabled

2. Stop the Gateway

openclaw gateway stop

MANDATORY before editing openclaw.json or cron/jobs.json. The gateway actively writes to jobs.json (updating job state after each cron run). Editing while the gateway runs causes race conditions and data loss.

3. Backup Config

cp ~/.openclaw/openclaw.json ~/.openclaw/openclaw.json.backup.$(date +%Y%m%d%H%M%S)

4. Create Directories

mkdir -p ~/.openclaw/workspace-<agent_id>/memory
mkdir -p ~/.openclaw/agents/<agent_id>/agent

NEVER reuse agentDir across agents — causes auth/session collisions.

5. Write Workspace Files

Use templates from assets/templates/ as starting points. Every agent needs:

SOUL.md is the most important file. Be specific about responsibilities. Include behavioral modes if the agent operates differently in different contexts (e.g., briefing mode vs chat mode).

6. Edit openclaw.json — Agent Entry

Add to agents.list[]. See references/config-schema.md for all valid fields.

Minimal entry:

{
  "id": "<agent_id>",
  "name": "<Display Name>",
  "workspace": "/home/archit/.openclaw/workspace-<agent_id>",
  "agentDir": "/home/archit/.openclaw/agents/<agent_id>/agent",
  "identity": { "name": "<Display Name>" }
}

Common additions:

• "model" — Override the default model cascade. Exclude expensive models for worker agents.
• "heartbeat": { "every": "0" } — Disable heartbeat for cron-only agents.
• "groupChat": { "mentionPatterns": ["@<id>", "@<Name>"] } — Enable @mentions in groups.

Only ONE agent should have "default": true (currently Fossil). The default agent receives all unrouted messages.

7. Edit openclaw.json — Telegram Routing (if needed)

THREE separate config changes are required. Missing any one causes silent failures. See references/telegram-routing.md for the full explanation.

1. Group config in channels.telegram.groups:

   "-100XXXXXXXXXX": { "requireMention": false }

1. Binding in bindings[]:

   { "agentId": "<id>", "match": { "channel": "telegram", "peer": { "kind": "group", "id": "-100XXXXXXXXXX" } } }

1. Mention patterns on the agent entry (already done in step 6 if groupChat was added).

8. Create Cron Jobs (if needed)

Edit cron/jobs.json. Every cron job prompt MUST include:

• Dynamic group ID resolution preamble (NEVER hardcode Telegram group IDs):

  FIRST: Resolve your Telegram group ID by running:
  jq -r '.bindings[] | select(.agentId == "<agent_id>") | .match.peer.id' ~/.openclaw/openclaw.json
  Use the output as the target for all Telegram messages in this task.

• Date injection: $(date '+%A, %B %d, %Y') after the preamble
• Explicit constraints: source allowlists, recency rules, format templates
• Delivery instructions: use target='<AGENT_GROUP_ID>' placeholder (resolved by the preamble)

This self-healing pattern ensures cron jobs survive Telegram group ID migrations. See references/prompt-patterns.md for full patterns and references/telegram-routing.md for why this matters.

Critical: If copying files or prompts from another agent's workspace, grep for hardcoded paths and update them.

9. Restart Gateway and Verify

openclaw gateway start

Verify in logs:

• Agent registered: agent registered: <id>
• Messages route correctly: lane enqueue: lane=session:agent:<id>:...

If messages to a Telegram group show skip: no-mention, the channels.telegram.groups config is missing (see references/bugs-and-pitfalls.md).

Reference Files

Template Files

Starter templates for workspace files are in assets/templates/. Copy and customize per agent.