Overview
This forum post shares the paper Phronix: The Capability-Centric Learning Paradigm for Agents (arXiv: 2610.31789) by Hariharan Devarajan, Neeraj Rajesh, Aniruddha Marathe, Nikoli Dryden, Kin Ng, and Michela Taufer. Field: machine learning. Posted 2026-10-27.
Key Points
- Problem: Scientific computing increasingly relies on LLM-based agents for performance engineering tasks (instrumentation, analysis, workflow porting), but agent effectiveness depends heavily on the surrounding harness, which typically relies on manually written skills and curated memory. This limits autonomous learning and forces agents to repeatedly rediscover solutions instead of retaining and reusing knowledge from prior executions.
- Approach: Phronix is a capability-centric learning paradigm enabling agentic harnesses to acquire, validate, and refine reusable capabilities from their own executions. It operates through a predictor-validator-corrector loop:
- MCP tools and associated skills propose solutions
- Tool executions validate them
- Observed outcomes refine stored capabilities
- Every observed interaction is distilled into a capability store; on task failure, Phronix investigates rather than abandons the attempt.
- Findings:
- Acquired capabilities solve repeated tasks with 93-95% first-attempt success across models, reducing solution non-determinism and improving reliability.
- Phronix converts an agent's own executions into 7.6-30.8 persistent capabilities per task without user guidance; capability creation drops to fewer than two when the needed capability already exists.
- Warm-start runs consume only 0.38-0.40 million tokens per task, 3-4x fewer than a single baseline attempt; cold-start costs do not exceed the baseline.
- On a 1000-genome Pegasus workflow and production application I/O profiling instrumentation tasks, Phronix passed all 11 correctness checks (baseline harness passed only 6).
- Phronix raised SWE-bench task resolution rates of two cheaper models from 29% and 27% to 91% and 87%, matching the most expensive evaluated model.
- Paper: https://arxiv.org/abs/2610.31789