Stop babysitting brittle agents.
One-off scripts rot the week they ship. SMITH treats agent building as engineering: declare the target, measure the miss, patch the architecture, prove the gain.
Agents that improve themselves
The forge loop rewrites architecture from real failures, not guesses. Generation 7 beats generation 1 by design.
Measured, never vibes
Accuracy, reliability, cost and latency are recorded on every run and compared side by side.
Domain agnostic
Same forge, different eval pack. Invoices today, repo claim checks tomorrow, your workflow next.
Failure taxonomy built in
Every miss lands in a class with a suggested patch, so fixes target the cause instead of the symptom.
Replayable traces
Full session traces for every generation. Reproduce any result months later.
Honest report cards
No cherry-picked demo. Thresholds are declared upfront and failures stay visible.
See a real before/after report card
Two domains, same forge, every generation scored.