AIAny
Icon for item

SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harness

Finds token- and API-cost-saving harness mechanisms for long-horizon coding agents using automated recursive self-improvement; packages four surviving mechanisms (action fusion, context compaction, observation archiving, delegated reading) to cut recorded token traffic ~44.7–49.0% and API cost by about one third while preserving most capability.

Introduction

Long-running coding agents repeatedly replay context, rerun verification steps, and reread large tool outputs — most savings come from changing the harness lifecycle, not the model. SoL‑Pi treats harness improvement as a searchable, recursive self-improvement problem: generate many candidate harness changes, pre-screen them, and validate survivors across diverse executable environments so improvements transfer beyond development settings.

Key Findings
  • Empirical gains: On a 51-task EdgeBench suite, SoL‑Pi retains roughly 94% of Pi's average score while reducing recorded token traffic by about 44.7–49.0% and cutting API cost by roughly one third.
  • Mechanisms discovered: Four narrowly scoped, composable mechanisms survived the automated search — Action Fusion (combine edit+test into one tool call), Online Context Compact (semantic compaction at subtask boundaries), ObservationPack (archive large payloads with lightweight handles and paged recall), and Evidence-Preserving Reducer (delegate first-pass log reading to cheaper agents with verifiable receipts).
  • Cost implications: Measured savings translate to estimated hourly cost reductions for a professional researcher on the evaluated backends (ranges reported per model/backend), demonstrating practical token- and cost-efficiency without large capability regressions.
Who it's for & trade-offs

Great fit if you run or research long-horizon coding agents and care about reducing API/token spend without substantially lowering task completion or evidence retention. SoL‑Pi is useful when harness-level changes are allowed and you can integrate optional mechanisms into Pi-style harnesses. Look elsewhere if your primary bottleneck is model quality rather than harness overhead, if you cannot modify the harness, or if strict reproducibility requires every tool output to be replayed into model context (some mechanisms archive content and recall pages on demand).

Where it fits

SoL‑Pi is a harness-level efficiency layer, not a new agent model. It complements improvements in base LLMs and retrieval stacks by reducing redundant context and model round-trips. Use it when operational cost and token throughput matter for continuous or unattended agent runs.

Mechanisms (brief)
  • Action Fusion: collapse predictable edit→test sequences into a single tool invocation to avoid an extra model round-trip.
  • Online Context Compact: trigger context compaction at semantic subtask boundaries when future savings justify the rewrite cost.
  • ObservationPack: replace repeated full-text replays with stable handles and paged recall, keeping exact content available without replaying it into every request.
  • Evidence-Preserving Reducer: delegate initial log scans to cheaper agents and bind verified receipts to archived logs so frontier agents only see quoted, verified lines.

Information

  • Websitearxiv.org
  • AuthorsHaozhe Liu, Tian Ye, Sensen Gao, Qihang Cao, Yitong Li, Mingchen Zhuge, Duomin Wang, Ruihua Zhang, Ping Luo, Jiawang Bian
  • Published date2026/09/17

Categories

More Items

Benchmarks joint recovery of obfuscated platform messages and evidence-grounded web investigations by pairing 3,600 synthetic restoration inputs with 600 controlled web environments. Separately scores message-entry recovery and downstream web risk judgments, and provides a frozen multimodal evidence judge plus a resettable local sandbox for reproducible agent evaluation.

Measures how individual harness components—planning, action space, and context management—affect coding agents' success, cost, and behavior. Uses a modular harness across 176 matched settings on SWE-Bench Verified and Terminal-Bench 2.1 to isolate component effects and surface model- and budget-dependent trade-offs.

Introduces physical rank consistency (PRC) and ActionPiece, a learned action tokenizer that preserves local physical relationships via joint supervision of representation learning and quantization, improving autoregressive vision-language-action policy performance across multiple benchmarks.