Agent S4 evolves its own interfaces, beats GPT-5.5 on OSWorld at 4.3% of the cost
Saaket Agashe · hf · 2026-10-21
The paper 'Evolving Agent Interfaces' introduces Agent S4, an autonomous computer-use scientist that evolves its own agent interfaces from task experience.
Key idea: computer-use agents are limited by human-designed interfaces—an agent may fill spreadsheet cells one by one when a single line of code would do. The authors argue interface design should itself be an agent capability:
- A meta-agent examines task agents' trajectories on unlabeled synthetic tasks, constructing and revising a repository of interfaces and usage knowledge
- Population-based evolution selects repositories by validation performance on tasks the meta-agent never sees, favoring generalizable changes
- On OSWorld V2, Agent S4 with Qwen3.8-27B sets a new open-source SOTA: 21.3% binary score, 3.8x the same model with fixed interfaces, above GPT-5.5 (13.0%) and Claude Opus 4.8 (20.6%) at 4.3% and 1.5% of their costs
- On ScienceBoard, evolved interfaces lift the same model's success rate from 41.4% to 58.6%
More from coding & agent
- OpenAI: Two Internal Coding Agents Weakened Their Own Misalignment Monitoring — rohanpaul_ai · 2026-10-21
- Tokyo DevDay: strict enterprise plugin policies push devs toward private plugins — sierracatalina · 2026-10-21
- New agent stack: Grok Bot orchestrates Gemini, ChatGPT and Claude via CLIs — _AustinCalvert_ · 2026-10-21
- Dev builds attest: position-aware prompt-injection scanner for MCP tool descriptions — mrpkk · 2026-10-21
- S-Agent 2.0 turns coding agents into spatial reasoners, hitting 66.5 on MMSI-Bench — liuziwei7 · 2026-10-21
- LLM trading agent skips 380 of 392 calls — scoring refusals shows skipped events moved 2x random — Tasty-Mammoth5321 · 2026-10-21