What an agent recipe actually looks like (and why it beats re-prompting every run)

Tej Pandya, founder of GrowEasy.ai

A "recipe" is the difference between a generic agent harness and a vertical one. Concretely, it is a codified procedure a specialist wrote once, which the harness executes every run instead of improvising:

recipe: qualify-inbound-call
steps:
  - load: caller_history from crm.byPhone(caller.id)
  - extract: {budget, timeline, intent} from transcript
  - check: budget >= pricing.floor          # deterministic
  - route: intent == "high" ? human_closer : nurture_sequence
verify:
  - schema: crm_entry.v2.json               # pinned output format
  - lint: no_unsupported_claims(summary)

Two parts of this are genuinely deterministic, and one is not. The routing code and the schema check are code - they run the same every time. The extract step calls a model, and served model output can vary even at temperature 0 (Thinking Machines, "Defeating Nondeterminism in LLM Inference"). So the honest claim is not "deterministic output". It is "predictable, verifiable outcomes": deterministic checks wrapped around a stochastic model. Martin Fowler's harness-engineering framing says the same - a harness increases the probability the agent gets it right.

Does it save tokens? Plausible - no re-explaining the task, fewer retries - but the recipe itself is injected context, and nobody has published a controlled with/without comparison that I could find. Anthropic reports agents use ~4x the tokens of chat and multi-agent setups ~15x, so architecture dominates token spend. Direction reasonable; magnitude unmeasured.

Does specialist procedure beat improvisation? The closest evidence: SWE-agent moved from 3.8% (a non-interactive retrieval baseline) to 12.47% of SWE-bench tasks on the same model purely through harness/interface design, and Voyager's verified skill library produced 3.3x more discoveries than recipe-less baselines in Minecraft - adjacent evidence, not a business benchmark.

One correction worth making: "generic" does not mean "no recipes". A general harness can load specialist skills. The real axis is recipes versus no recipes, not product category. Recipes rot and cover only known cases; generic wins on novelty. Write the recipe after you've done the work the hard way.

Sources: arxiv.org/abs/2405.15793 (SWE-agent), arxiv.org/abs/2305.16291 (Voyager), anthropic.com/engineering/built-multi-agent-research-system, thinkingmachines.ai/blog/defeating-nondeterminism-in-llm-inference, martinfowler.com/articles/harness-engineering.html