autonomous research agent — best-first experiment tree
tree search (budget $30)
seed / root
hypothesis + config
sp top-4
sp top-8 (*)
sp top-16
lr=3e-4
sp8 + lr=1e-3
sp8 + dropout
sp8 + layer-drop
chosen branch
novelty x quality
expand() proposes k children by small-edit variation
each child runs in a sandboxed container:
docker run --network=none --memory=8g --cpus=2 --read-only
verify() checks: loss converged? baselines ran? seed set?
score() = 0.4*novelty + 0.5*quality + 0.1*remaining_budget
frontier is a max-heap; pop highest score each step
writer loop
LaTeX draft v1
vision critique
Opus 4.7 reads PDF
draft v2, v3...
reviewer ensemble
5 judges, NeurIPS rubric
safety
sandbox per experiment
no network egress
bounded wall-clock
deterministic seeds
cgroup memory + pids caps
$30 hard budget / paper
red-team: fork bombs,
filesystem escape,
LLM shell-metachar injection
Sakana v2 documented attack surface