AI Briefing — June 5th, 2026

🔬 Research

ArcANE: Do Role-Playing Language Agents Stay in Character at the Right Time?

34 upvotes · Song et al. · Paper

This benchmark tests whether role-playing language models can evolve their character's psychology as a story progresses, rather than maintaining fixed personas. ArcANE spans 17 novels and 80 characters, testing responses both within and outside the source text. The research shows that conditioning on "Character Arcs" beats other context strategies across six models, with the largest gains on scenarios outside the source text where retrieval methods find nothing.

TIDE: Proactive Multi-Problem Discovery via Template-Guided Iteration

30 upvotes · Jeong et al. · Paper

TIDE addresses a key limitation in AI agents: they only respond to explicit user requests, missing hidden problems in documents or code. The framework uses iterative discovery to find multiple coexisting issues and thought templates to ground predictions in recognizable problem classes. Tests on personal workspaces and software repositories show substantial improvements over single-shot approaches in coverage and resolution.

VideoKR: Towards Knowledge- and Reasoning-Intensive Video Understanding

25 upvotes · Fu et al. · Paper

VideoKR introduces the first large-scale training corpus specifically for knowledge-intensive video reasoning, with 315K examples across 145K expert-domain videos. The human-in-the-loop pipeline targets progressively deeper reasoning while ensuring reliability. Models trained on VideoKR outperform existing approaches on knowledge-intensive video tasks while staying competitive on general video reasoning, highlighting data design as crucial for progress.

🛠 Tools

NousResearch/hermes-agent — The agent that grows with you

1913/day · Repo

Hermes Agent is a self-improving AI agent with a built-in learning loop that creates skills from experience and builds a deepening model of users across sessions. It runs anywhere from a $5 VPS to GPU clusters, supports 200+ models through various providers, and can be accessed remotely via Telegram. The agent's key differentiator is persistent learning and memory that spans conversations and improves over time.

chopratejas/headroom — Compress tool outputs, logs, files, and RAG chunks

3142/day · Repo

Headroom tackles the token cost problem by compressing various inputs before they reach LLMs, achieving 60-95% token reduction while maintaining answer quality. It works as a library, proxy, or MCP server, addressing a critical cost optimization need as AI tool usage scales up. Given Uber's recent $1,500/month AI tool caps, compression tools like this could become essential for enterprise adoption.

affaan-m/ECC — Agent harness performance optimization system

1750/day · Repo

ECC is a comprehensive agent harness system designed for Claude Code, Codex, and other coding agents, featuring skills, memory, security, and research-first development. With 207K+ stars and extensive multi-language support, it positions itself as infrastructure for optimizing agentic coding workflows. The system appears to focus on making coding agents more efficient and secure in production environments.

📡 Analysis

AI enthusiasts are in a race against time, AI skeptics are in a race against entropy

June 4th · Post · tags: ai, charity-majors, agentic-engineering

Charity Majors identifies a critical organizational challenge: AI enthusiasts see existential threats from competitors using AI to ship faster, while skeptics worry about reliability degradation from shipping code faster than engineers can understand it. Both perspectives are valid, creating a leadership challenge around designing feedback loops to bridge the "gap in shared reality" between these groups within the same teams.

Uber Caps Usage of AI Tools Like Claude Code to Manage Costs

June 3rd · Post · tags: ai, llm-pricing, coding-agents, uber

Uber implements $1,500 monthly spending caps per AI coding tool per employee after blowing through their 2026 AI budget in four months. This represents roughly 11% of median engineer compensation ($330K), suggesting Uber sees significant value in these tools. The cap allows for substantial usage while preventing runaway costs, offering a practical model for enterprise AI adoption at scale.

Sources: HuggingFace Papers API, GitHub Trending, simonwillison.net