Applied AI
Agent architecture without the hype
By the end you'll know when to use a single agent versus a multi-agent system, what observability requires, and the failure modes that don't show up in toy demos.
4 steps · ~20 minutes of reading total
- 1
The agent architecture decision
EmpiricaThe single-vs-multi-agent decision, framed around failure tolerance, observability, and whether subtasks are actually independent.
- 2
Milestone: you can name three failure modes only multi-agent systems have
MilestoneIf you can't name them, you don't yet know whether your design is single or multi by accident.
- 3
Anthropic — Building effective agents
Anthropic ↗First-principles framing on when to use workflows vs agents. Worth reading in full before committing to either.
- 4
The real cost of LLM API calls
EmpiricaAgent loops multiply token usage in ways that aren't obvious from a small demo. Architecture choices have direct cost consequences.