Five-agent team structure

Context and Problem Statement

The previous setup had 15+ scattered agents with overlapping roles, unclear ownership, and expensive model usage. Agents were being spawned as disposable subagents, losing context between runs. We needed a lean, permanent team with clear roles and cost-efficient model assignments.

Decision

Consolidate to five core agents with fixed roles and dedicated models:

Rules:

• Never spawn core agents as disposable subagents — use sessions_send for inter-agent communication
• Each agent has persistent sessions and memory
• Johnny orchestrates and delegates, never writes code
• T-Bug coordinates CLI tools only, never touches code directly
• All crons run on cost-efficient models (MiniMax/Kimi), not Anthropic

Consequences

• Good, because clear ownership eliminates role confusion and duplicate work
• Good, because 5 agents with persistent sessions retain context vs 15 disposable ones that don’t
• Good, because model assignments match role requirements (expensive models only where needed)
• Bad, because fewer agents means less parallelism — but focused work beats scattered effort
• Neutral, because the team can be expanded later if a clear gap emerges

Implementation Plan

Affected files

• openclaw.json → agents.list array
• MEMORY.md → team roster section
• IDENTITY.md → role descriptions
• All cron scripts → model references

Verification

• openclaw status shows exactly 6 agents (5 core + codex ACP)
• No orphaned agent sessions from old roster
• Inter-agent messaging works via sessions_send