Multi-agent: patterns when one agent is not enough
Multi-agent systems: router, specialists, critic, orchestrator, subagents (Isolate). Only if single-agent is unreadable. Token and hop governance.
One LLM is not an architecture
Gluing an LLM behind an API is not multi-agent design. Agentic architecture combines reasoning, tools, memory, control policies and observability.
Useful patterns: intent router, domain specialist, deterministic tool, state memory, critic/reflex, orchestrator.
Switch to multi-agent when monster prompts and tools step on each other, not for LinkedIn prestige.
In SMEs: stabilize one agent + HITL first (how to create an AI agent article).
Isolate: subagents and separate contexts
A subagent has its own context window. The orchestrator does not swallow the full research history. That is Isolate from context engineering applied to architecture.
Useful for: research vs writing vs execution; headless CI jobs (plan → patch → test) if isolation and secrets are serious.
Without max hops, timeouts and circuit breakers, agents talk in loops and burn tokens.
Multi-agent governance
Each agent: owner, token budget, tool allowlist (MCP), metrics. The orchestrator has a written delegation policy.
HITL at risk boundaries (customer send, financial write), not only “at the end”.
Cost: multi-agent often means multiple LLM calls. Budget and cache. Not a license to waste.
Roadmap
1) Stabilize one useful single agent. 2) Extract a second specialist. 3) Add a router. 4) Measure. 5) Orchestrate only if needed.
If you want to validate whether multi-agent is justified, we can scope in 20-40 minutes.
FAQ
- Smarter than one agent?
- Not automatically. More modular, often more expensive and fragile when poorly designed.
- How many agents at the start?
- One, then two. Six agents on day one is an anti-pattern.
- Need a multi-agent framework?
- Only if single agent + clear code is not enough. A framework without observability adds opaque magic.
- Link to MCP?
- MCP standardizes tools. The more agents, the more allowlists and schemas matter.
- Link to context engineering?
- Isolate and Compress are the main levers in multi-agent. See the dedicated article.
Scope your first AI agent
20 minutes to review your tools, data and the first useful case. No jargon, no commitment.