Build an agent that reads, writes, and tests code
one move: close the tool loop safely
Open the code pane immediately, then use the Spec tab, optional design question, tests, and artifact when you want them.
Finish each move before carrying the artifact forward.
An agent is a model-driven control loop constrained by typed tools, observations, and a sandbox.
Deliverable: Write the mechanism in one sentence.Implement JSON tool calls, a ReAct loop, file operations, test execution, and stop conditions.
Deliverable: Type the smallest runnable implementation.Tool schemas, path boundaries, turn budgets, and failure recovery must be explicit.
Deliverable: Record the expected output and one edge case.Break tools, inject malformed plans, exceed budgets, and compare recovery policies.
Deliverable: Change one variable and explain the result.A traced coding-agent run with a sandbox and safety policy.
Deliverable: Save the code, output, and a short failure note.Tool schemas, path boundaries, turn budgets, and failure recovery must be explicit.
Break tools, inject malformed plans, exceed budgets, and compare recovery policies.
A traced coding-agent run with a sandbox and safety policy.
Mapped from the supplied Build Everything PDF, source page 63. The five moves are Deriva’s implementation contract for this project.