Labs
Three in-browser experiments
Implementations of the agent loop, the context budget, and an evaluation harness. The policy can be replaced later; the trace, the budget, and the scorers are the object of study.
Lab 01
Tool-calling agent loop
Think → tool call → observation → answer. The policy here is deterministic so every hop is inspectable. Replace the policy with a model later; keep the protocol.
tools: retrieve, calc
Lab 02
Context budget
The window is a memory hierarchy. Allocate tokens the way a kernel allocates RAM. When you overflow, compact. Pin the contract.
Lab 03
Eval harness
An eval is a frozen contract over model behavior. Edit the candidate outputs, run the scorers, watch pass rate move. This is the regression suite for non-deterministic code.
-
pass rate