The anatomy of an agent
Understand the rule, inspect a real Level Higher build, make the decision yourself and leave with a reusable artifact.
An agent is not a clever chat window. It is a small operating system: a goal, approved context, bounded tools and evidence that tells it when to stop.
If one part stays implicit, fluent output can hide an unfinished job.
Name the change
Describe what must exist when the work is complete.
Bound the system
List the context and tools the agent may use - and what it must never touch.
Define proof
Choose evidence another person can inspect without trusting the agent.
Turn “blur faces” into a verifiable local workflow.
A vague request could upload private video, miss moving faces or export an unusable file.
The build contract fixed five things: local-only processing, the input video, face detection and tracking tools, an export target and a frame-level verification pass.
The operator can inspect what entered the system, what authority it had and what proves the privacy-preserving export exists.
Select a node to inspect what fails when it is missing.
Map one real agent before adding tools.
Replace the example with one task you genuinely want to finish. Keep every line observable.
- The objective describes a finished state
- Unknowns are named instead of guessed
- Tool authority is smaller than the entire system
- Evidence can be checked independently
Edit the worksheet here. It stays in your browser until you download or copy it.
A useful agent loop is inspectable from objective to evidence.
Your work should now pass the four checks above. If one fails, repair the contract or evidence before adding more tools.