AP
Agentic Playbook
OpenClaw·Intermediate·Last tested: 2026-03·~10 min read

Agent Roles

A well-designed multi-agent system assigns each agent a clear purpose, model, and tool set. Below are example agent roles that cover common organizational functions.

Coordinator Agent

Purpose: The primary command-center agent. Morning briefs, KPI tracking, cross-functional summaries, strategic analysis.

Recommended model: anthropic/claude-opus-4-6 (complex reasoning, synthesis across domains)

Channel bindings:

  • DMs from the primary administrator
  • Catch-all default for unmatched messages

Tool access:

  • read, write, exec, web_search, memory_search
  • sessions_spawn (to delegate tasks to sub-agents)
  • sessions_list, sessions_send (to query other agents' outputs)

Workspace files:

  • SOUL.md — Strategic advisor persona with deep knowledge of your business
  • USER.md — Admin preferences, communication style, timezone
  • AGENTS.md — Morning brief protocol, KPI thresholds, alert rules
  • HEARTBEAT.md — Periodic checks (e.g., every 30 min: check dashboards, flag anomalies)

Key behavior: This agent can spawn sub-agents on cheaper models for parallel research (e.g., "analyze last week's churn data" + "summarize competitor news" simultaneously).

Engineering Agent

Purpose: Code reviews, PR checks, documentation, GitHub issue management, CI/CD monitoring, technical troubleshooting.

Recommended model: anthropic/claude-sonnet-4-6 (fast, good at code, cost-effective for high-volume tasks)

Channel bindings:

  • #engineering channel
  • #devops channel
  • GitHub webhooks (PR notifications, CI failures)

Tool access:

  • read, write, exec, process, git
  • web_search (for documentation lookup)
  • Deny: sessions_spawn, gateway, cron (no spawning, no gateway control)

Sandbox: Run in Docker sandbox — this agent executes code and should be isolated from the host.

Workspace files:

  • SOUL.md — Senior developer persona, knows your tech stack
  • AGENTS.md — PR review standards, coding conventions, commit message format
  • TOOLS.md — References to repositories, CI pipeline URLs, database schema catalog

Product Agent

Purpose: Write specs, user stories, documentation, competitive analysis, translate product requirements between business and engineering.

Recommended model: anthropic/claude-sonnet-4-6

Channel bindings:

  • #product channel
  • #general (with mention pattern @product-agent)

Tool access:

  • read, write, web_search, memory_search
  • Deny: exec, process, git, gateway (no code execution needed)

Workspace files:

  • SOUL.md — Product manager persona with domain expertise
  • AGENTS.md — Feature spec templates, user story format, prioritization framework
  • TOOLS.md — Documentation platform API access, project management references

Risk / Analytics Agent

Purpose: Data analysis, financial modeling, portfolio monitoring, regulatory reporting support.

Recommended model: anthropic/claude-opus-4-6 (needs strong reasoning for analysis)

Channel bindings:

  • #analytics or #risk channel
  • DMs from analytics team members only (via peer allowlist)

Tool access:

  • read, write, web_search
  • READ-ONLY database access — this agent queries but never writes to production
  • Deny: exec (use a custom skill wrapper for SQL queries with read-only connection)
Read-Only Enforcement

Always wrap database access for analytics agents in a custom skill with a read-only connection string. Never give direct write access to production databases.

Workspace files:

  • SOUL.md — Analyst persona with domain-specific regulatory knowledge
  • AGENTS.md — Assessment protocols, scoring model parameters, alert thresholds
  • TOOLS.md — Database connection details (read-only), API endpoints

Sales Agent (Phase 2)

Purpose: Answer sales questions, generate client reports, track pipeline, analyze customer relationships.

Recommended model: anthropic/claude-sonnet-4-6

Channel bindings:

  • #sales channel

Tool access:

  • read, write, web_search, memory_search
  • Deny: exec, process, gateway
Tip

Deploy this agent after your first three or four agents are stable. Adding agents incrementally reduces coordination complexity.

Model Selection Guidelines

| Agent Type | Recommended Model | Rationale | |-----------|-------------------|-----------| | Coordinator / Executive | Opus | Complex reasoning, cross-domain synthesis | | Engineering | Sonnet | Fast, strong at code, cost-effective for volume | | Product / Documentation | Sonnet | Good writing quality at lower cost | | Risk / Analytics | Opus | Financial reasoning, regulatory analysis | | Sales / Support | Sonnet | Conversational, good enough for Q&A | | Sub-agents (spawned) | Sonnet | Cost control — the parent agent does synthesis |