most agent frameworks fix the strategy at design time. when it fails it fails the same way forever, and you are the one who has to notice.
this keeps a lightweight reinforcement learning bandit over its own approaches, so tactics that keep working get chosen more often.
a small idea tested honestly rather than a framework. the interesting result was how fast it learned to stop doing the thing that never worked.
the loop
planner, implementer, reviewer. the planner emits a json plan that says how it should be validated. the implementer writes python with tests, using openai when it is available and templates when it is not. the reviewer runs it through pytest.
the bandit
an epsilon greedy bandit picks between three modes, fast, safe and detailed, and keeps reward statistics for each. every episode is logged with its task, mode, reward and reflection.
questions that are not code get graded by the model against a rubric, from 0 to 1, with 0.6 counting as success. a failure writes a reflection the next plan can read.
