BlackArc18 papers
Agent Architecture & Harnesses
The loop, the harness, and the plumbing that turns a model into a working agent.
- Papers
- 18
- Read time
- 4.9h
- Avg length
- 16m
ReAct: Synergizing Reasoning and Acting in Language Models
Let an LLM interleave free-text "thoughts" with real tool actions in a single prompt loop, and it plans, grounds itself in the world, and recovers from mistakes far better than reasoning-only or acting-only agents — this is the paper that defined the agent loop you already use.
- Read
- 16 min
- Source
- arXiv:2210.03629
Agent Harness Engineering: A Survey
The reliability of an AI agent depends as much on the plumbing wrapped around the model — sandboxes, tools, memory, control loops, tracing, evals, and guardrails — as on the model itself, and this survey names that plumbing "the harness," gives it a seven-part map (ETCLOVG), and grades 170+ real projects against it.
- Read
- 22 min
- Source
- arXiv:2507.15330
Dive into Claude Code: The Design Space of Today's and Future AI Agent Systems
A line-by-line reverse-engineering of Claude Code's TypeScript source that shows a production coding agent is not a clever model loop — it is a thin reasoning core (~1.6% of the code) wrapped in a dense deterministic harness for permissions, context compaction, extensibility, delegation, and persistence, and that harness is where all the real engineering (and your build/moat opportunity) lives.
- Read
- 22 min
- Source
- arXiv:2604.14228
A Practical Guide for Designing, Developing, and Deploying Production-Grade Agentic AI Workflows
A field-tested checklist of nine engineering rules that turn flaky LLM-agent demos into deterministic, observable, deployable production systems — distilled from building a real news-to-podcast pipeline.
- Read
- 16 min
- Source
- arXiv:2512.08769
A Practical Approach for Building Production-Grade Conversational Agents with Workflow Graphs
Instead of stuffing every business rule into one giant system prompt, Kakao splits their e-commerce agent into a directed graph of small, single-purpose LLM states — then fine-tunes on the resulting conversations with a masking trick that stops each state's training data from contaminating the others.
- Read
- 15 min
- Source
- arXiv:2505.23006
From Agent Behaviour to Agent-Friendly Documentation - An Empirical Study of How Coding Agents Discover, Read, and Write Technical Documentation
When you actually watch 557 real coding-agent sessions instead of guessing, agents spend 60.5% of their documentation time on files written *for* agents (AGENTS.md, CLAUDE.md, plans, `thoughts/`) and only 1.3% on API references — and almost none of the "read the docs, then write the code, then check the docs" behaviour everyone assumes actually shows up in the traces.
- Read
- 22 min
- Source
- arXiv:2608.20195
JIT-Agent: Scaling Harness Intelligence via Just-in-Time Harness Evolution
JIT-Agent is a trained "meta-agent" that, given a task, writes a bespoke agent
- Read
- 15 min
- Source
- arXiv:2608.25593
Meta-Harness: End-to-End Optimization of Model Harnesses
Meta-Harness points a coding agent at a folder full of every previous harness's source code, scores, and raw execution logs, and lets it rewrite the harness itself — instead of hand-tuning prompts, it automates the harness-engineering job directly.
- Read
- 14 min
- Source
- arXiv:2603.28052
Model or Harness? An Interaction-Centric Taxonomy for Localizing Agent Failures
When your agent fails, this paper gives you a fixed vocabulary for saying *which two parts of the system were talking to each other* and *which of the two was actually wrong* — so the fix lands on the model, the scaffolding, the environment, or the test, instead of on whatever you guessed.
- Read
- 18 min
- Source
- arXiv:2607.28802
OpenJarvis: Personal AI, On Personal Devices
Swapping a cloud model for a local one breaks a personal AI assistant because the whole stack was built around the cloud model, not just the model; OpenJarvis turns the entire stack into an editable, searchable object so a cloud "teacher" can rebuild it around the local model instead.
- Read
- 14 min
- Source
- arXiv:2605.17172
Self-Refine: Iterative Refinement with Self-Feedback
Same LLM, three hats — write a draft, critique your own draft, rewrite it — looped a few times, and you get 5-40% better outputs with zero training, fine-tuning, or extra models.
- Read
- 14 min
- Source
- arXiv:2303.17651
Tendem: A Hybrid AI+Human Platform
Tendem runs an AI agent through a plan-act-observe-verify loop and only hands work to a paid human expert at specific "step gates" where risk or uncertainty is high, and on 94 real freelance-style tasks that combination beats both a pure AI agent and pure human freelancers on quality, speed, and cost at once.
- Read
- 9 min
- Source
- arXiv:2602.01119
The Scaffolding Matters More Than the Interface
They ran the same six-step GitHub task through seven agent harnesses, five models, and both tool interfaces, and found that which harness you pick changes your bill by up to 139×, while MCP-versus-CLI does not reliably change it at all.
- Read
- 22 min
- Source
- arXiv:2608.08654
Toolformer: Language Models Can Teach Themselves to Use Tools
A 6.7B model teaches itself, with no human labels, exactly when to call a calculator, search engine, translator, QA system, or calendar mid-sentence — by generating its own API calls, keeping only the ones that measurably make its next-word predictions easier, and fine-tuning on that self-curated dataset.
- Read
- 10 min
- Source
- arXiv:2302.04761
Toward Training Superintelligent Software Agents through Self-Play SWE-RL
A single coding LLM teaches itself to fix bugs by also learning to *create* bugs — playing both attacker and defender against itself inside real codebases — and gets steadily better at solving real GitHub issues it was never trained on.
- Read
- 22 min
- Source
- arXiv:2512.18552
Voyager: An Open-Ended Embodied Agent with Large Language Models
Voyager is a Minecraft agent that never stops playing — it picks its own next goal, writes JavaScript to pursue it, keeps every script that worked in a growing skill library, and uses that library to learn faster and faster, all without ever touching GPT-4's weights.
- Read
- 14 min
- Source
- arXiv:2305.16291
Web World Models
Instead of letting an LLM hallucinate an entire game/world, define the world's rules and state in ordinary web code and let the model only fill in the flavor — giving you infinite, controllable environments with no database.
- Read
- 16 min
- Source
- arXiv:2512.23676
WebGPT: Browser-assisted question-answering with human feedback
OpenAI gave GPT-3 a text-only web browser and a "quote" button, then used human preference data to train it to search, click, and cite its way to answers that beat human-written Reddit answers 69% of the time.
- Read
- 14 min
- Source
- arXiv:2112.09332