PhyAgentOS Benchmarking Architecture

Benchmarking is an Agent workflow skill that discovers tasks, creates standard Runtime Sessions, and aggregates immutable execution evidence.

中文

1 · Boundary

Benchmarking Agent Skill

Selects an enabled simulation Target and policy Skill Runtime, discovers tasks, appends Sessions, aggregates evidence, and writes a report.

Policy Skill Runtime

Executes one policy closed loop through SessionRunner, Target, Policy Client, Adapter, and Bridge. It does not orchestrate experiments.

Benchmarking does not create a new SKILLRUNTIME entry and does not instantiate SessionRunner directly.

2 · Availability Gate

runtime enabled → enabled simulation Target → supported policy Skill Runtime → skill visible

SkillsLoader checks the Runtime workspace and target-enable overrides. This controls skill visibility; Runtime Preflight still validates every Session.

3 · Lifecycle

Load protocol → Select Target/Policy pair → Discover tasks → Append pending Sessions → Watchdog execution → ResultWriter → Aggregate report

4 · Evidence

SESSIONS.md

Final status, success, steps, return, latency, artifact path, and error code.

LOG.md

Session-indexed execution history.

episode.json

Benchmark metadata, task/instance, chunks, reward, steps, and result metadata.

5 · Report Contract

Reports identify setup, suite/task source, metrics, missing-value policy, aggregate and per-task results, failure taxonomy, artifact paths, reproduction settings, and comparison scope. Reports reference evidence without rewriting it.

6 · Source Map

agent/skills.py · agent/context.py · skills/benchmarking/SKILL.md · runtime/watchdog/supervisor.py · runtime/artifacts/episode_writer.py