Agent Patterns

Agent Patterns

Canonical recipes live in AgentLab.md. This page is a pointer + a quick index.

Solo patterns

• 24/7 autonomous SDR — agent that watches inbound, ingests, qualifies, escalates
• Autonomous AE — deal-owner that moves stages on real events
• Research agent — nightly enrichment
• Morning briefing agent — daily standup from the timeline

Team / swarm patterns

• Outbound campaign machine — planner + executor + analyst
• End-to-end deal orchestration — researcher + negotiator + recorder
• Inbound triage + routing swarm — specialty routing via tag-claimed tasks
• Territory-aware swarm — agents scoped by geography
• Post-call follow-up swarm — transcript → notes → tasks

The ten rules

From the nakatomi-crm Claude skill:

1. Start with describe_schema.
2. Never paste the API key into the conversation.
3. Reads before writes (get_*, timeline).
4. Upsert with external_id.
5. Use the relationship graph (relate).
6. Log activities generously.
7. Soft delete by default.
8. Idempotency for critical writes.
9. Don't invent fields — use data.
10. Link memories after writing rich context.

Harness recipes

Claude Desktop, Cursor, Claude Code, ChatGPT Custom GPTs, Perplexity, Hermes / OpenClaw — each has a snippet in AgentLab → Harness recipes.

Connector chains

How to stack Nakatomi with the tools agents already have:

• Nakatomi + Gmail / Hostinger
• Nakatomi + Google Calendar
• Nakatomi + Slack
• Nakatomi + Apollo / HubSpot / ImportYeti
• Nakatomi + memory systems

Contributing a pattern

If you've built an agent pattern in production and it works, PR it into AgentLab.md. Good recipes have:

• A one-line goal
• The trigger (cron, webhook, human prompt)
• The actual tool calls in order
• What Nakatomi does vs. what external tools do
• One anti-pattern you hit while building it

Keep each pattern under ~60 lines.