nova
The agent that does the work. Everything else in laserbrain measures work, manages it or records it — nova is the one that acts. It holds skills, runs fleets, and is measured against a ground it has no way to move.
pip install laserbrain
what it is
from laserbrain import Nova
n = Nova(goal='ship the parser and the benchmark')
n.learn('search', my_search_fn)
n.run(act) # act(ctx) -> {goal, progress, distance, done?}The thinking arrives as act — usually a model. nova supplies the loop, the skills and the instrumentation. It is a scaffold, and saying so is the point: the part that matters is that the loop takes a reading every step and there is no flag to skip it.
skills
the one thing it cannot do
Move the ground it is measured against.
nova has no method that sets, moves or clears its ground — the harness freezes it at the first check. An agent that can revise the reference it is measured against is measuring itself, which is the one thing the proof rules out.
Python offers no true barrier, so nova offers evidence instead: the ground is fingerprinted at the first reading and ground_intact() answers directly. Tampering does not raise — it is reported. A monitor that crashes gets removed; one that tells you gets read.
a fleet
out = n.compose({'a': agent_a, 'b': agent_b, 'c': agent_c})
out['_nova']['seen_only_from_above'] # 1
out['_nova']['collisions'] # [{'agents': ['a','b'], 'overlap': 0.8}]Two agents handed the same job are both perfectly on their own ground and correct at every step. The duplication exists only as a relation, so it is invisible from inside either one however capable it is. That is what a second vantage buys — not better thinking, a different position.