kengdevLet’s talk
Updated on 2026-09-10
A multi-agent system splits a problem across several agents, each with its own context, tools and scope, coordinated by an orchestrator. You get there when a single agent starts confusing unrelated tasks, not because the architecture looks more elegant.
Recognise when splitting genuinely helps, and when it only adds failure modes.
Two agents cost more than twice one agent, in tokens and in debugging. Split only when one is genuinely stuck.
One LLM is not an architecture
Gluing an LLM behind an API does not make a multi-agent design. A real agentic architecture combines reasoning, tools, memory, control policies and observability.
Six patterns come back in the systems that hold up in production:
| Pattern | Role |
|---|---|
| Intent router | It routes each request to the right specialist |
| Domain specialist | It handles a bounded scope with its own tools |
| Deterministic tool | It makes computation and critical actions reliable |
| State memory | It keeps progress between steps |
| Critic / reflex | It rereads and fixes the output before it ships |
| Orchestrator | It sequences the agents and enforces the policy |
You switch to multi-agent when prompts become monstrous and tools step on each other, not for LinkedIn prestige.
In an SME, stabilize a single agent with its HITL first (see the how to create an AI agent article).
Isolate: subagents and separate contexts
A subagent works in its own context window. The orchestrator therefore does not swallow the full research history. That is the Isolate lever from context engineering, applied to architecture.
The split becomes useful when research, writing and execution each deserve their own context, or for headless CI jobs that chain the plan, the patch and the test, as long as you take isolation and secrets seriously.
Without a hop limit, timeouts and circuit breakers, agents answer each other in loops and burn tokens.
Multi-agent governance
Each agent has an owner, a token budget, a tool allowlist (MCP) and its metrics. The orchestrator enforces a written delegation policy.
Place HITL at the risky boundaries, like a customer send or a financial write, not only "at the end".
On cost, a multi-agent system often multiplies LLM calls. Budget, add caching, and keep in mind that the architecture is no license to waste.
Roadmap
The reasonable progression fits in five steps:
- Stabilize one useful single agent
- Extract a second specialist
- Add a router
- Measure
- Orchestrate, only if needed
If you want to validate that a move to multi-agent is justified, we can scope it in 20-40 minutes.
Frequently asked questions
Smarter than one agent?
A multi-agent system is not automatically smarter. It is more modular, often more expensive, and it turns fragile when poorly designed.
How many agents at the start?
Start with one agent, then move to two. Starting with six agents on day one is an anti-pattern.
Need a multi-agent framework?
A framework is only justified once a single agent and clear code are no longer enough. Without observability, it mostly adds opaque magic.
Link to MCP?
MCP standardizes the tools. The more agents you run, the more allowlists and schemas matter.
Link to context engineering?
Isolate and Compress are the two main context engineering levers in multi-agent. The dedicated article details them.
Sources and references
- Building effective agentsAnthropic
Primary documentation for product capabilities and changes.
