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.

Multimodal large language models (MLLMs) have demonstrated strong performance across diverse multimodal tasks, achieving promising outcomes. However, their application to emotion recognition in natural images remains underexplored. MLLMs struggle to handle ambiguous emotional expressions and implicit affective cues, whose capability is crucial for affective understanding but largely overlooked. To address these challenges, we propose MERMAID, a novel multi-agent framework that integrates a multi-perspective self-reflection module, an emotion-guided visual augmentation module, and a cross-modal verification module. These components enable agents to interact across modalities and reinforce subtle emotional semantics, thereby enhancing emotion recognition and supporting autonomous performance. Extensive experiments show that MERMAID outperforms existing methods, achieving absolute accuracy gains of 8.70%–27.90% across diverse benchmarks and exhibiting greater robustness in emotionally diverse scenarios.