Papers and technical reports behind the 2026 agent engineering stack.
Reasoning models & test-time compute
| Paper / report |
Why it matters |
Link |
| DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning |
Open reasoning model — RL-trained chain-of-thought |
arXiv:2501.12948 |
| Let's Verify Step by Step (Process Reward Models) |
Reward each reasoning step, not just the final answer |
arXiv:2305.20050 |
| Scaling LLM Test-Time Compute Optimally |
How much inference compute to spend per task |
arXiv:2408.03314 |
| OpenAI o1 System Card |
Production reasoning model design choices |
OpenAI PDF |
Context & memory engineering
| Paper |
Why it matters |
Link |
| Lost in the Middle: How Language Models Use Long Contexts |
Why context window placement matters — center gets ignored |
arXiv:2307.03172 |
| MemGPT: Towards LLMs as Operating Systems |
Virtual memory paging for long conversations |
arXiv:2310.08560 |
| Large Language Models as Tool Makers |
Agents that create reusable tools/skills |
arXiv:2305.17126 |
Skills, programs & automated optimization
| Paper |
Why it matters |
Link |
| Voyager: An Open-Ended Embodied Agent with Large Language Models |
Growing skill library from experience |
arXiv:2305.16291 |
| DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines |
Optimize prompts/programs automatically |
arXiv:2310.03714 |
| TextGrad: Automatic Differentiation via Text |
Gradient-free optimization of agent pipelines |
arXiv:2406.07496 |
Coding agents
| Paper |
Why it matters |
Link |
| SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering |
ACI design — foundation for coding agents like Claude Code |
arXiv:2405.15703 |
| SWE-bench: Can Language Models Resolve Real-World GitHub Issues? |
The benchmark coding agents are measured against |
arXiv:2310.06770 |
| Devin / Cognition technical blog |
End-to-end software engineering agent architecture |
cognition.ai |
Harness & safety
| Paper |
Why it matters |
Link |
| Constitutional AI: Harmlessness from AI Feedback |
Self-critique loops for safer agents |
arXiv:2212.08073 |
| AgentPoison: Red-teaming LLM Agents via Poisoning Memory |
Why memory and tool outputs need sanitization |
arXiv:2407.12784 |
Protocols & standards (not papers, but essential)
| Resource |
What it is |
Link |
| Model Context Protocol (MCP) |
Standard for tool/resource servers |
modelcontextprotocol.io |
| OpenTelemetry GenAI Semantic Conventions |
Standard spans for LLM/agent tracing |
OTel docs |
Also see