Course 2 - Module 4 - Lesson 1
Research, coding, support, and operations agents as repeatable recipes
Useful Agent Recipes
A recipe connects input to evidence
An agent recipe describes the trigger, allowed information, actions, checks, and finish. It should be understandable before anyone implements it. Start with a specific output, such as a source-backed answer draft, rather than a broad instruction to manage an entire department.
A research recipe might gather official sources, extract relevant passages, compare claims, and produce a brief with links and unanswered questions. Its check is whether each important claim has support. Its finish includes unresolved gaps, not a promise that every possible source was searched.
Compare three other recipes
A coding recipe begins with a reproducible problem, proposes a limited change, runs appropriate checks, and presents the difference for review. Evidence includes the observed failure and the check after the change. Generating code is one step, not the whole definition of success.
A support recipe reads an approved policy and drafts a reply grounded in that policy. An operations recipe compares a stock report with reorder thresholds and flags exceptions. Support should escalate an absent policy; operations should flag missing counts rather than silently treating them as zero.
Make the review match the job
Different recipes need different reviewers. A researcher checks sources, a maintainer checks software behavior, and a support owner checks promises to customers. Define who takes over when the agent reaches uncertainty beyond its instructions or a tool returns an unexpected result.
Test each recipe using a normal example, an incomplete example, and an example outside its scope. Watch where the steps diverge. These trials help you improve a reusable method while keeping the permissions and responsibility attached to the actual job.
Tiny experiment
Choose research, coding, support, or operations. Write a recipe with one input, two actions, a measurable check, and a clear escalation condition.
Confidence check
What evidence would your reviewer need to distinguish a completed job from a convincing completion message?