Agent Control Flow

The order planned actions execute — sequential, parallel, conditional, or looped — with the model itself often determining non-sequential branches.

Control flow is the order in which planned actions run. Simple agents stay sequential: task B waits on task A (for example, SQL runs only after NL-to- SQL translation). Richer flows include:

• Parallel — independent calls run together (fetch prices for many products concurrently), cutting user-perceived latency.
• Conditionals — branch on prior output (buy vs. sell after reading an earnings report).
• Loops — repeat until a condition holds.

Unlike traditional software, AI agents often let the model decide control flow. That makes non-sequential plans harder to generate and harder to translate into executable commands (especially with natural language action plans). When choosing an agent framework, check which control flows it supports — parallel browse/fetch is a common latency win that sequential-only harnesses cannot express.

Fixed workflow topology (pipeline, DAG, optional cycles) is the non-model-driven counterpart: routing is code, while individual tasks may still use LLMs. Prefer that shape until goals need LLM-driven workflow routing.

---

Sources

• AI Engineering (Chip Huyen), ch. 6
• Prompt Engineering for LLMs (Berryman), ch. 9