Kev's Dream Team (14+ Agents)

14+ agents under one gateway with Opus 4.5 orchestrator delegating to Codex workers. Comprehensive technical write-up covering the Dream Team roster, model selection, sandboxing, webhooks, heartbeats, and delegation flows.

Kev's Dream Team by @adam91holt is one of the most ambitious multi-agent architectures documented in the OpenClaw community. It runs 14+ specialized agents under a single OpenClaw gateway, with an Opus 4.5 orchestrator (named 'Kev') that delegates tasks to purpose-built workers including Codex-powered coders, Gemini Flash scouts for research, Opus analysts for deep thinking, and Gemini Pro agents for visual tasks. The technical write-up is comprehensive and opinionated. It covers the full architecture: how the gateway routes between agents, model selection rationale for each specialist, sandboxing strategies to prevent agents from interfering with each other, webhook integrations for external triggers, heartbeat patterns for proactive behavior, and delegation flows that determine which agent handles which task. The manifesto argues that 'Orchestration > Chat' — the future isn't talking to one AI, it's having a coordinator that manages a team of specialists. The project is accompanied by a public GitHub repo with the articles, architecture diagrams (Kev at the top delegating to Rex, Scout, Hawk, Pixel, and others), and the manifesto document. It's become a reference architecture for anyone wanting to build sophisticated multi-agent setups with OpenClaw, and it highlights how different models have different strengths that can be composed into a more capable whole.

Tags: multi-agent, orchestration, architecture, manifesto

Category: devtools

Tips

  • Start with 2-3 specialized agents before scaling to 14+ — get the delegation patterns right with a small team first
  • Use different models for different strengths: fast models (Gemini Flash) for research/triage, powerful models (Opus) for analysis, coding models (Codex) for development
  • Implement clear naming conventions for your agents — the Dream Team uses character names (Kev, Rex, Scout, Hawk, Pixel) to make delegation flows readable
  • Use sandboxing to prevent agents from stepping on each other's work — especially important when multiple coding agents operate on the same codebase
  • Document your architecture as thoroughly as @adam91holt did — multi-agent setups become unmaintainable without clear documentation of who does what

Community Feedback

@adam91holt took things further with something called 'Kev's Dream Team.' It's a gateway system running 14+ specialised agents, all orchestrated by an Opus 4.5 coordinator. One brain directing a small army of AI workers.

— Generative AI Publication

Our principles: Orchestration > Chat, Specialization > Generalization, and the future of Deployable Intelligence Units.

— GitHub

14+ agents under one gateway with Opus 4.5 orchestrator delegating to Codex workers. Comprehensive technical write-up covering the Dream Team roster, model selection, sandboxing, webhooks, heartbeats, and delegation flows.

— OpenClaw Showcase

Frequently Asked Questions

How much does running 14+ agents cost?

Costs depend heavily on which models you use and how active each agent is. The orchestrator (Opus 4.5) handles routing decisions, while cheaper models (Gemini Flash) handle routine tasks. Expect $50-200/month for an active multi-agent setup.

Do all 14 agents run simultaneously?

No. The orchestrator spawns specialist agents on demand. Most agents are dormant until the coordinator determines their expertise is needed. Only the orchestrator runs continuously.

Can I replicate this architecture?

Yes. The GitHub repo includes the manifesto and technical write-up with architecture details. The key components are OpenClaw's sub-agent spawning, model routing, and heartbeat features, all available in standard OpenClaw.

What's the 'Deployable Intelligence Unit' concept?

It's @adam91holt's framework for thinking about agent teams as deployable units — self-contained groups of specialists that can be versioned, tested, and deployed together, similar to how microservices teams work in software engineering.