Every AI agent I've shipped follows one of three patterns. Same three patterns that run every human organization.
- Single Agent:
One LLM, one prompt, a set of tools (search, APIs, calculators). You define the goal, not the steps. The model's internal reasoning figures out the sequence on its own. "Plan a trip to San Francisco" and it autonomously searches weather, books flights, finds hotels.
It's just like a solo developer: flexible, fast to set up. But non-deterministic. Ask it to find a late-night restaurant AND calculate the fastest route, and the prompt becomes fragile. It skips steps, calls tools in the wrong order, hallucinates. Maximum flexibility, zero control.
- Sequential Agent:
A fixed pipeline where agents pass the baton through a shared session state, a short-term memory scratchpad. Agent 1 finds restaurants, writes to the scratchpad. Agent 2 reads it, maps the route. The system guarantees food is found before transport is planned.
Like an assembly line: predictable, reliable. But rigid. If the user needs step 3 before step 1, it can't adapt.
- Parallel Agent:
Independent agents running simultaneously. Museum finder, concert finder, restaurant finder, all at once. A final aggregator agent synthesizes everything into one cohesive output.
Imagine a cross-functional team with a project manager. 3x faster. But more expensive, and engineering the "gather and synthesize" step is harder than it looks.
From what I see, Claude Code uses parallel sub-agents for codebase research. Cursor routes different models per task type. Your daily tools already picked one of these patterns.