Natural Language Action Plans

Express agent plans in natural language then translate to tool calls, which is more robust to API churn than planning directly in exact function names.

Planning with exact function names is brittle. Tool APIs rename and reshape over time, forcing prompt and few-shot updates — and a finetuned planner would need refinetuning. Cross-use-case reuse also suffers when each environment has different APIs.

An alternative is to generate plans in natural language ("get current date / retrieve best-selling product last week / …"). Models trained mostly on natural language handle this better and hallucinate less on the plan itself. The cost is a translator (sometimes called a program generator) that converts each NL action into an executable command from the tool inventory.

Translation is usually simpler than planning and can run on a weaker model with lower hallucination risk — but the translator itself becomes a failure point and needs its own checks under agent planning failure modes and agent tool failure modes.

---

Sources

• AI Engineering (Chip Huyen), ch. 6