When it pays off
A single agent works well while the task fits in one context and one way of thinking. When the work mixes research, writing, review and decision, each step needs different instructions, tools and criteria. Putting it all in one agent creates a large, confusing context, which brings in the problems described in context engineering.
The three patterns
- Sequence. Each agent does one step and hands off to the next. In Branding OS, six agents go from brief to finished piece, and a validator checks the brand before human approval.
- Parallel. Several agents work at the same time on the same input, without seeing each other's results. In experiment #003, two parallel auditors found 16 issues one alone would have missed.
- Orchestrator. One agent plans, splits the work and hands it to specialists. In ATHENA OS, a request goes through 7 phases and is executed by 9 types of agent, each with its own context.
Where these systems fail
- At the handoff. If an agent does not know exactly what it receives and what it must deliver, the error spreads to the next steps.
- In shared context. When everyone reads everything, one agent's error contaminates the others. That is why each agent should have its own context.
- In too many agents. More agents does not mean more power. In experiment #007, the rule that worked was "agents think, skills execute": fewer agents, each one more complete.
Frequently asked questions
How many agents does a system need?
The minimum that separates steps with different demands. If two agents use the same context and the same criteria, they are probably one.
Is a multi-agent system more expensive?
Not necessarily. Simple steps can go to smaller models. In the measured case, 71.76% of the work was done by sub-agents.
How do I know the final result is right?
Put a check between steps, not only at the end. The guide how to know whether an AI answer is right shows which kind fits each step.
Want to apply this to your operation?
The first call is a free one-hour diagnosis.