The Archive

ORPilot: open-source agentic system translating ambiguous business problems into solver-ready optimization models with conversational and data collection agents.

Zero-trust authorization framework for LLM agents with hybrid inspection and task-based access control to mitigate tool-use and resource-access risks.

Novel framework for bandwidth-efficient remote control via minimal information transmission between controller and agents in continuous action spaces.

DataEvolver implements closed-loop agent-driven visual data generation and refinement for image editing, supporting masks, depth, poses, and trajectory artifacts.

Momentum integrates runtime procedural content generation and autonomous agent evaluation in endless-runner gameplay to assess generated terrain balance and solvability.

Iterated negotiation benchmark tests LLM agents' ability to repair grounding failures in dynamic multi-turn interaction.

GRAVITY module injects relational, temporal, and thematic structure into conversational memory retrieval for long-horizon agents.

AutoMat benchmark evaluates LLM agents on reproducing computational materials science findings, requiring domain knowledge and result interpretation beyond code quality.

MemCoE: cognition-inspired two-stage memory optimization for LLM agents to learn personalized long-term user preferences within context windows.

Analysis of agentic AI specialization: coding agents (Codex-style) for knowledge work, Claude for creative tasks; discusses agents escaping operational boundaries.

Intern-Atlas introduces structured methodological evolution graphs as research infrastructure for AI agents to navigate scientific knowledge beyond citation links.

Link lets users connect cards, banks, and subscriptions, then authorize AI agents to spend securely via approval flows.

STEF enables schema-agnostic evaluation of text-to-SQL agents in production without ground-truth queries, addressing real-world deployment gaps.

Study finds persona prompting in multimodal LLMs produces stable but limited behavioral variation in urban sentiment judgment tasks.

CARE methodology systematizes LLM agent engineering in scientific domains via three-party collaboration between SMEs, developers, and helper agents.

NVIDIA CUDA Tile (cuTile) is a tile-based programming model that enables developers to write GPU kernels in terms of tile-level operations—loads, stores, and... NVIDIA CUDA Tile (cuTile) is a tile-based programming model that enables developers to write GPU kernels in terms of tile-level operations—loads, stores, and matrix multiply-accumulate—rather than manually coordinating threads, warps, and shared memory. cuTile.jl brings the same tile-based approach to the dynamic programming language Julia. Users can write custom GPU kernels without dropping… Source

Architectural pattern language for vision language agents balances latency/non-determinism of VLMs against real-time enterprise control requirements.

Comparative evaluation of three LLM agent paradigms (domain-specific, computer-use, coding) on scientific visualization tasks across 15 benchmarks.

D3-Gym dataset: 565 verifiable tasks from real scientific repositories for evaluating LLM agents on data-driven discovery.

Language model agents design mechanical linkages via symbolic lifting: LLMs explore topologies while numerical optimizers tune parameters, validated on six motion targets.

Survey of RL+GUI agents for long-horizon automation in visual interfaces; proposes framework toward autonomous digital inhabitants with safe exploration.

Schema-grounded external memory for agents outperforms text-retrieval approaches by enabling exact fact tracking, state updates, and structured queries.

Survey formalizing graph-based world models for agents, decomposing environments into entity nodes and edges to improve robustness vs. flat-tensor approaches.

On-demand persona-based agent generation framework enabling dynamic multi-agent workflow customization without hard-coded architectures.

Lightweight clinical agent architecture using integrated state dynamics to surface pre-escalation risk signals in LLM clinical deployment.

KellyBench: long-horizon sequential decision benchmark using 2023-24 Premier League sports betting; evaluates agents on non-stationary open-ended optimization.

Working on large codebases with Claude Code, we kept running into the same issue: when Claude looks for relevant code, it falls back to grep, reading full files, or launching multiple subagents. This burns through tokens, and often misses the relevant code. There are some existing solutions (that we also benchmarked against), but they all had issues (too slow, needs API keys, quality not good enough, etc). We built [Semble](https://github.com/MinishLab/semble) to fix this. It's a local MCP server that gives Claude Code high quality code search: instead of reading files to find what's relevan...

tl;dr: your skill in AI is a measure of your **quality** and **scale**. Use **success criteria** and **subagents** intentionally to get excellent results. Use skills and .md docs when you find repeating patterns in your daily work, not before. **---** **Quality** comes from telling the agent what outcome you want, and the **success criteria** that you will use to measure a “good” outcome. This helps avoid Claude's tendency to rush completion. Note this is specifically *not* telling it what to *do*, but instead what to *achieve*. If you come from the old world, you might remember terms like ...

I can have multiple, dense legal documents on my screen, each 40, 60, or 100+ pages each with the Claude Word add-in agents syncing, pushing and pulling information between them, pinging each other, and providing helpful context so that I can draft all three or four in parallel or ensure that an entire package is consistent. I can have a lengthy spreadsheet workbook open containing 10 worksheets and the information is analyzed and pulled in by the agents when needed. I am absolutely blown away at how well this is implemented and the improvement in quality, consistency and efficiency. It ...

ClawGym framework for systematic development of file/tool-based agents, includes 13.5K synthesized tasks for agent training and evaluation.