Start from the job.

The canon: agent-computer interfaces, SWE-bench as the yardstick, and the frameworks that survived contact with real repos.

• arxiv:2310.06770 — SWE-bench: can LLMs resolve real GitHub issues?
• github:princeton-nlp/SWE-agent — SWE-agent — the agent-computer interface
• github:aider-AI/aider — aider — pair programming in the terminal
• github:AllHandsAI/OpenHands — OpenHands — open platform for software agents
• arxiv:2305.16291 — Voyager — lifelong-learning agent with a skill library

From Toolformer's learned API calls to the Model Context Protocol's standard interface — typed, callable surfaces.

• arxiv:2210.03629 — ReAct — reasoning and acting in a loop
• arxiv:2302.04761 — Toolformer — models teaching themselves to call APIs
• github:modelcontextprotocol/modelcontextprotocol — Model Context Protocol — the standard tool interface
• arxiv:2307.16789 — ToolLLM — mastering real-world API spaces

Roles, conversations, and SOPs: when one agent is not enough.

• github:microsoft/autogen — AutoGen — multi-agent conversation framework
• github:crewAIInc/crewAI — crewAI — role-playing multi-agent framework
• arxiv:2304.03442 — Generative Agents — interactive simulacra of human behavior
• arxiv:2308.10848 — AgentVerse — dynamic multi-agent collaboration

Benchmarks, harnesses, tracing: measurement before vibes.

• arxiv:2310.06770 — SWE-bench — real-issue evaluation
• arxiv:2308.03688 — AgentBench — evaluating LLMs as agents
• github:openai/evals — openai/evals — canonical eval harness examples
• github:Arize-ai/phoenix — Phoenix — LLM observability and eval tracing
• github:langfuse/langfuse — Langfuse — traces and evals

Paged context, reflection, hippocampus-inspired retrieval — the whole memory stack.

• arxiv:2310.08560 — MemGPT — LLMs as operating systems
• github:letta-ai/letta — Letta — stateful agents (MemGPT successor)
• github:mem0ai/mem0 — mem0 — memory layer for AI apps
• arxiv:2405.14831 — HippoRAG — long-term memory for LLMs
• arxiv:2303.11366 — Reflexion — verbal reinforcement learning

RAG from first principles: the original paper, the lost-in-the-middle caveat, and the engines.

• arxiv:2005.11401 — RAG — retrieval-augmented generation (the original)
• arxiv:2307.03172 — Lost in the Middle — position effects in long contexts
• github:qdrant/qdrant — qdrant — vector similarity engine
• github:microsoft/graphrag — GraphRAG — graph-based retrieval

Chains, trees, and search over thoughts — deliberate problem solving.

• arxiv:2201.11903 — Chain-of-thought prompting
• arxiv:2305.10601 — Tree of Thoughts — deliberate problem solving
• arxiv:2310.04306 — LATS — language agent tree search
• github:stanford-oval/storm — STORM — grounded long-form writing system

Guaranteed structure: grammar constraints, validated schemas, compiled pipelines.

• github:dottxt-ai/outlines — outlines — guaranteed-structure generation
• github:guidance-ai/guidance — guidance — constrained generation language
• arxiv:2310.03714 — DSPy — compiling declarative LLM pipelines
• github:pydantic/pydantic-ai — pydantic-ai — typed, validated agent outputs

Local and high-throughput inference, and the library underneath.

• github:vllm-project/vllm — vLLM — high-throughput inference engine
• github:ollama/ollama — ollama — run open models locally
• github:huggingface/transformers — transformers — the canonical open-model library