Scientific computing increasingly relies on large language model agents for tasks such as instrumentation, analysis, and workflow porting. The effectiveness of these agents, however, is tightly bound to manually crafted skills and curated memory, forcing them to rediscover solutions repeatedly and limiting autonomous learning.
Phronix introduces a capability‑centric learning paradigm that lets an agent's harness acquire, validate, and refine reusable capabilities from its own executions. The core is a predictor‑validator‑corrector loop: MCP tools and associated skills propose solutions, tool executions validate them, and observed outcomes update the capability store. Every interaction is distilled into a capability entry, and failing tasks are investigated rather than abandoned.
Results show that learned capabilities solve recurring tasks on the first try in 93%‑95% of cases, dramatically reducing nondeterminism and improving reliability. Each task yields 7.6‑30.8 durable capabilities on average, and once a needed capability exists, creation drops below two. Warm‑start runs consume only 0.38‑0.40 M tokens per task—three to four times less than a single baseline attempt—while cold‑start costs do not exceed the baseline.
On the 1000 Genome Pegasus workflow for I/O performance analysis, Phronix passes all 11 correctness checks compared with six for the baseline harness, and lifts two cheaper models' task‑solving rates from 29% and 27% to 91% and 87%, matching the performance of the most expensive evaluated model.
Review