The 2026 edition · 119 breakdowns
The Agent Atlas
Every important paper on how AI agents actually get built, read closely, distilled to what a builder can use, and mapped into one navigable atlas. Architecture, memory, context, retrieval, reasoning, security, and the research the whole field stands on.
Browse by topics
- 01 Agent Architecture & Harnesses 6
- 02 Memory Systems 9
- 03 Context Engineering 8
- 04 Multi-Agent Systems 7
- 05 Reasoning & Test-Time Compute 7
- 06 Self-Improving Agents 12
- 07 Retrieval & RAG 5
- 08 Knowledge Graphs 6
- 09 Security & Safety 15
- 10 Evaluation & Benchmarks 4
- 11 Reinforcement Learning 6
- 12 Applied & Industry 18
- 13 Foundations & Infrastructure 16
01 Featured this edition
The papers that define how agents get built
Chosen for how much they shaped real agent engineering: the loop, the harness, memory, context, and the security that keeps it standing.
Browse every topic →
Agent Architecture & Harnesses ReAct Synergizing Reasoning and Acting in Language Models 2023 · 16 min · arXiv 2210.03629 
Self-Improving Agents Reflexion Language Agents with Verbal Reinforcement Learning 2023 · 16 min · arXiv 2303.11366 
Agent Architecture & Harnesses Agent Harness Engineering A Survey 2026 · 22 min · arXiv 2507.15330 
Context Engineering A Survey of Context Engineering for Large Language Models 2025 · 22 min · arXiv 2507.13334 
Memory Systems HippoRAG Neurobiologically Inspired Long-Term Memory for Large Language Models 2024 · 16 min · arXiv 2405.14831 
Memory Systems MemGPT Towards LLMs as Operating Systems 2024 · 14 min · arXiv 2310.08560 
Security & Safety Design Patterns for Securing LLM Agents against Prompt Injections 2025 · 22 min · arXiv 2506.08837 
Multi-Agent Systems Why Do Multi-Agent LLM Systems Fail? 2025 · 16 min - 01 ReAct 16
- 02 Reflexion 16
- 03 Agent Harness Engineering 22
- 04 A Survey of Context Engineering for Large Language Models 22
- 05 A Survey of Context Engineering for Large Language Models 22
- 06 Fundamentals of Building Autonomous LLM Agents 16
- 07 Agentic Context Engineering 18
- 08 HippoRAG 16
Start here if you read one thing
Agent Architecture & Harnesses · 2023
ReAct
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.
Browse
Explore by topic
Agent Architecture & Harnesses
The loop, the harness, and the plumbing that turns a model into a working agent.
- Papers
- 6
- Read time
- 1.9h
Memory Systems
How agents retain, recall, and reflect across long-running work.
- Papers
- 9
- Read time
- 2.3h
Context Engineering
Packing the right tokens into the window: retrieval, compression, and ultra-long context.
- Papers
- 8
- Read time
- 2.6h
Multi-Agent Systems
Orchestration, collaboration, roles, and why teams of agents succeed or fail.
- Papers
- 7
- Read time
- 1.9h
Reasoning & Test-Time Compute
Chains, trees, atoms, and System-2 thinking that trade compute for correctness.
- Papers
- 7
- Read time
- 2.0h
Self-Improving Agents
Agents that evolve their own prompts, skills, and harnesses from experience.
- Papers
- 12
- Read time
- 3.8h
Retrieval & RAG
Grounding generation in external knowledge, from classic RAG to reasoning retrieval.
- Papers
- 5
- Read time
- 1.4h
Knowledge Graphs
Graph-structured memory and retrieval for agents that need to connect facts.
- Papers
- 6
- Read time
- 1.7h
Security & Safety
Prompt injection, access control, guardrails, and the risk surface of autonomous agents.
- Papers
- 15
- Read time
- 4.3h
Evaluation & Benchmarks
Judging agents honestly: benchmarks, LLM-as-judge, and outcome measurement.
- Papers
- 4
- Read time
- 1.2h
Reinforcement Learning
Training agents end-to-end with rewards, from GRPO to deployment loops.
- Papers
- 6
- Read time
- 1.7h
Applied & Industry
Agents in the real world: supply chain, operations research, business process, discovery.
- Papers
- 18
- Read time
- 5.2h
Foundations & Infrastructure
The substrate: LLM internals, GPUs, serving systems, and the theory underneath.
- Papers
- 16
- Read time
- 5.3h
Most foundational
The papers everything else builds on
BlackArc Industrial
This atlas is a byproduct of the work.
BlackArc builds production AI agents for manufacturing and operations. The Agent Atlas is how we stay current: every paper here feeds what we ship. If you are putting agents into real operations, come talk to us.
Method
How this atlas was built
Read the papers, not the abstracts
Every entry is a close reading of one paper: the problem, what is genuinely new, how it works down to the math, and how you would actually build with it.
Distill to what a builder can use
Each breakdown ends where most summaries stop: pseudocode of the core idea, how to make your own version, and where it breaks.
Map it into one publication
119 breakdowns, sorted into 13 topics, so the whole shape of agent research is browsable in one place.