@inproceedings{zhou-etal-2026-0xclaw,
title = "0x{C}law: Autonomous Open-ended Agents with Phase-aware Budget Allocation and Temporal Pressure",
author = "Zhou, Yu and
Dou, Xinyuan and
Cao, Yudong and
Song, Junhao",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.415/",
pages = "8753--8769",
ISBN = "979-8-89176-462-0",
abstract = "Large language model agents handle structured coding tasks well, but open-ended projects are harder. A hackathon gives an agent an open goal, a hard deadline and several artifacts to deliver. The agent has to decide how to split its budget across phases and when to stop widening scope. Most agents do neither. 0xClaw is a multi-agent framework that splits a hackathon run into seven phases. Each phase hands a typed artifact to the next. The Budget Router is a small fine-tuned Qwen3.5-2B model. It reads the workspace state and predicts a per-phase control token for token budget, time, reasoning effort and repair rounds. The Deadline Framer turns each token, together with the cumulative execution log, into a plain deadline instruction given to the agent at the start of each phase. Our study covers 20 hackathon tasks, 10 baselines and more than 200 agent runs. 0xClaw matches Claude Code at the best pass@1 (1.00) and exceeds ARIS (0.95), the two strongest baselines. Against Claude Code, it achieves comparable quality and pass@1 while using 59{\%} fewer tokens. The trade-off is wall-clock time, about 2.5 times that of ARIS."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="zhou-etal-2026-0xclaw">
<titleInfo>
<title>0xClaw: Autonomous Open-ended Agents with Phase-aware Budget Allocation and Temporal Pressure</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yu</namePart>
<namePart type="family">Zhou</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xinyuan</namePart>
<namePart type="family">Dou</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yudong</namePart>
<namePart type="family">Cao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Junhao</namePart>
<namePart type="family">Song</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Large language model agents handle structured coding tasks well, but open-ended projects are harder. A hackathon gives an agent an open goal, a hard deadline and several artifacts to deliver. The agent has to decide how to split its budget across phases and when to stop widening scope. Most agents do neither. 0xClaw is a multi-agent framework that splits a hackathon run into seven phases. Each phase hands a typed artifact to the next. The Budget Router is a small fine-tuned Qwen3.5-2B model. It reads the workspace state and predicts a per-phase control token for token budget, time, reasoning effort and repair rounds. The Deadline Framer turns each token, together with the cumulative execution log, into a plain deadline instruction given to the agent at the start of each phase. Our study covers 20 hackathon tasks, 10 baselines and more than 200 agent runs. 0xClaw matches Claude Code at the best pass@1 (1.00) and exceeds ARIS (0.95), the two strongest baselines. Against Claude Code, it achieves comparable quality and pass@1 while using 59% fewer tokens. The trade-off is wall-clock time, about 2.5 times that of ARIS.</abstract>
<identifier type="citekey">zhou-etal-2026-0xclaw</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.415/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>8753</start>
<end>8769</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T 0xClaw: Autonomous Open-ended Agents with Phase-aware Budget Allocation and Temporal Pressure
%A Zhou, Yu
%A Dou, Xinyuan
%A Cao, Yudong
%A Song, Junhao
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F zhou-etal-2026-0xclaw
%X Large language model agents handle structured coding tasks well, but open-ended projects are harder. A hackathon gives an agent an open goal, a hard deadline and several artifacts to deliver. The agent has to decide how to split its budget across phases and when to stop widening scope. Most agents do neither. 0xClaw is a multi-agent framework that splits a hackathon run into seven phases. Each phase hands a typed artifact to the next. The Budget Router is a small fine-tuned Qwen3.5-2B model. It reads the workspace state and predicts a per-phase control token for token budget, time, reasoning effort and repair rounds. The Deadline Framer turns each token, together with the cumulative execution log, into a plain deadline instruction given to the agent at the start of each phase. Our study covers 20 hackathon tasks, 10 baselines and more than 200 agent runs. 0xClaw matches Claude Code at the best pass@1 (1.00) and exceeds ARIS (0.95), the two strongest baselines. Against Claude Code, it achieves comparable quality and pass@1 while using 59% fewer tokens. The trade-off is wall-clock time, about 2.5 times that of ARIS.
%U https://aclanthology.org/2026.findings-emnlp.415/
%P 8753-8769
Markdown (Informal)
[0xClaw: Autonomous Open-ended Agents with Phase-aware Budget Allocation and Temporal Pressure](https://aclanthology.org/2026.findings-emnlp.415/) (Zhou et al., Findings 2026)
ACL