Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Agentic Engineering Reference Library

This page is a verified index of primary sources for agentic engineering: official documentation, developer and API portals, source repositories, SDKs, downloadable or offline documentation, a two-track learning path, and free-access research literature.

It is a navigation layer, not a tutorial. Where the rest of this book explains a concept, this page tells you which document to open to get the authoritative answer, and in what order to read things. Every link was HTTP-verified on the date shown below; sources that block automated checkers but work in a browser are flagged rather than silently dropped.

Protocols (MCP, A2A), agent frameworks, coding agents, sandboxed execution and durable workflows, tracing and benchmarks, agent security, and a two-track path from reading a thousand-line agent loop to evaluating and isolating production agents.

71 entries across 7 categories, plus 54 education & reference-implementation resources (21 basic / 33 advanced), plus 9 video resources and 5 conference sources.

Every link HTTP-verified on 2026-10-07.

This is the fastest-moving topic in the set. Check publication dates on everything — framework APIs here break more often than anywhere else in this series.

Contents

1. Protocols & interoperability

Model Context Protocol (MCP)

  • Docs: modelcontextprotocol.io
  • Developer / API: modelcontextprotocol.io/specification/…
  • Source: github.com/modelcontextprotocol
  • SDKs & repos: SDKs for Python, TypeScript, Java, Kotlin, C#, Go, Rust; Inspector debugging tool
  • Downloadable / offline: Spec is a single readable document, versioned by date
  • Note: The emerging standard for connecting models to tools and data. Read the spec directly — it is short, and it explains the trust boundaries most integrations get wrong.

MCP reference servers

A2A (Agent2Agent)

  • Docs: a2a-protocol.org/latest
  • Source: github.com/a2aproject/A2A
  • SDKs & repos: Agent-to-agent communication: agent cards, task delegation, streaming
  • Downloadable / offline: Spec + SDKs
  • Note: Complements MCP rather than competing with it — MCP connects agents to tools, A2A connects agents to each other. Linux Foundation governed.

OpenAI function calling

  • Docs: platform.openai.com/docs/…
  • Developer / API: platform.openai.com/docs
  • SDKs & repos: The tool-call interface most frameworks ultimately target
  • Downloadable / offline: Docs site
  • Note: Learn the raw protocol before adopting a framework. Most agent bugs are tool-schema bugs.

OpenTelemetry GenAI semantic conventions

2. Agent frameworks

OpenAI Agents SDK

Claude Agent SDK

LangGraph

  • Docs: langchain-ai.github.io/langgraph
  • Source: github.com/langchain-ai/langgraph
  • SDKs & repos: Agents as explicit state graphs; checkpointing, human-in-the-loop, time travel
  • Downloadable / offline: Docs site
  • Note: The graph model is the right abstraction once control flow matters. Durable checkpointing is the feature that makes agents operable.

Pydantic AI

  • Docs: ai.pydantic.dev
  • Source: github.com/pydantic/pydantic-ai
  • SDKs & repos: Type-safe agents with validated structured output and dependency injection
  • Downloadable / offline: Docs site
  • Note: Feels like normal Python rather than a DSL. Strong choice if your codebase already uses Pydantic.

smolagents

CrewAI

AutoGen

  • Docs: microsoft.github.io/autogen
  • Source: github.com/microsoft/autogen
  • SDKs & repos: Microsoft’s multi-agent conversation framework; event-driven core
  • Downloadable / offline: Docs site
  • Note: The most research-driven of the multi-agent frameworks; papers accompany the design.

Semantic Kernel

Google Agent Development Kit

LlamaIndex Workflows

Mastra

  • Docs: mastra.ai/docs
  • Source: github.com/mastra-ai/mastra
  • SDKs & repos: TypeScript agent framework: workflows, RAG, evals
  • Downloadable / offline: Docs site
  • Note: The most complete option if your stack is TypeScript rather than Python.

3. Coding agents — read these, they are the most mature

Claude Code

  • Docs: docs.claude.com/en/docs/claude-code/overview
  • SDKs & repos: Terminal coding agent; hooks, subagents, MCP, skills
  • Downloadable / offline: Docs site
  • Note: Study the hooks and permissions model — it is a worked answer to ‘how do you let an agent act without letting it act freely’.

OpenHands

  • Docs: docs.all-hands.dev
  • Source: github.com/All-Hands-AI/OpenHands
  • SDKs & repos: Open-source software-development agent with a sandboxed runtime
  • Downloadable / offline: Docs site
  • Note: Fully open and genuinely capable. The runtime isolation design is the part to read.

SWE-agent

  • Docs: swe-agent.com/latest
  • Source: github.com/SWE-agent/SWE-agent
  • SDKs & repos: The Princeton research agent that established the agent-computer interface idea
  • Downloadable / offline: Docs site
  • Note: The ACI paper is the key insight: agent performance depends more on tool design than on prompting.

Aider

  • Docs: aider.chat/docs
  • Source: github.com/Aider-AI/aider
  • SDKs & repos: Terminal pair programmer with repository mapping and git integration
  • Downloadable / offline: Docs site
  • Note: The repo-map approach to context selection is clever and well documented.

Cline

Continue

browser-use

Playwright MCP

4. Runtime, sandboxing & durability

E2B

  • Docs: e2b.dev/docs
  • Source: github.com/e2b-dev/E2B
  • SDKs & repos: Firecracker-backed sandboxes for running agent-generated code
  • Downloadable / offline: Docs site; self-hostable
  • Note: If your agent executes code it wrote, it needs one of these. Not optional.
  • Docs: modal.com/docs
  • SDKs & repos: Serverless containers with fast cold starts; common agent execution backend
  • Downloadable / offline: Docs site

Temporal

LangGraph persistence

5. Observability, evaluation & benchmarks

Langfuse

  • Docs: langfuse.com/docs
  • Source: github.com/langfuse/langfuse
  • SDKs & repos: Open-source tracing for multi-step agent runs; self-hostable
  • Downloadable / offline: Docs site
  • Note: Agent debugging is trace debugging. Without this you are reading logs and guessing.

LangSmith

Phoenix (Arize)

OpenLLMetry

Inspect AI

SWE-bench

  • Docs: swebench.com
  • Source: github.com/SWE-bench/SWE-bench
  • SDKs & repos: Real GitHub issues from real repositories; the dominant coding-agent benchmark
  • Downloadable / offline: Harness and datasets in the repo
  • Note: Verified, Lite and Full are different benchmarks. Scores get quoted without saying which — always check.

Terminal-Bench

  • Docs: tbench.ai
  • SDKs & repos: Agents operating a real terminal on real tasks
  • Downloadable / offline: Free
  • Note: Less saturated than SWE-bench and harder to game.

Berkeley Function-Calling Leaderboard

  • Docs: gorilla.cs.berkeley.edu/leaderboard.html
  • SDKs & repos: Standardised tool- and function-calling evaluation
  • Downloadable / offline: Free
  • Note: The most useful single signal for whether a model can call tools reliably.

6. Guardrails, permissions & agent security

OWASP Top 10 for LLM Applications

  • Docs: genai.owasp.org
  • Developer / API: owasp.org/…
  • SDKs & repos: ‘Excessive agency’ and ‘insecure output handling’ are the agent-specific entries
  • Downloadable / offline: Free PDFs
  • Note: Agents turn prompt injection from an information-disclosure bug into a remote-code-execution bug. Read this first.

Simon Willison — the lethal trifecta

  • Docs: simonwillison.net/tags/prompt-injection
  • Developer / API: simonwillison.net
  • SDKs & repos: Private data + untrusted content + external communication = unsafe, regardless of prompting
  • Downloadable / offline: Free archive
  • Note: The single most useful design rule in agent engineering. Remove one leg of the trifecta.

Guardrails AI

NeMo Guardrails

garak

MITRE ATLAS

  • Docs: atlas.mitre.org
  • SDKs & repos: Adversarial taxonomy for AI systems
  • Downloadable / offline: Free
  • Note: Use it to structure an agent threat model.

7. Research papers & open-access literature

arXiv

  • Docs: arxiv.org
  • Developer / API: info.arxiv.org/help/api/index.html
  • SDKs & repos: Preprints across all of CS; most systems, ML and PL work appears here before publication
  • Downloadable / offline: Every paper is a free PDF. Bulk access documented at info.arxiv.org/help/bulk_data/index.html; a full-text API at export.arxiv.org
  • Note: Not peer-reviewed. Treat an arXiv-only paper as a claim, not a result — but it is where you will read almost everything first.

ar5iv

  • Docs: ar5iv.labs.arxiv.org
  • SDKs & repos: Renders any arXiv paper as responsive HTML instead of PDF
  • Downloadable / offline: Free; swap arxiv.org/abs/ID for ar5iv.labs.arxiv.org/html/ID
  • Note: Makes papers readable on a phone and searchable in-page. Underused.

alphaXiv

  • Docs: alphaxiv.org
  • SDKs & repos: arXiv papers with a public comment and discussion layer
  • Downloadable / offline: Free
  • Note: Useful when a paper is contested — the discussion often contains the critique you were looking for.

Semantic Scholar

  • Docs: semanticscholar.org
  • Developer / API: api.semanticscholar.org/graph/v1
  • SDKs & repos: 200M+ papers with citation graph, influential-citation scoring and TLDR summaries
  • Downloadable / offline: Free Graph API (semanticscholar.org/product/api), bulk datasets available on request
  • Note: The best free citation graph. ‘Highly influential citations’ is a genuinely useful filter for finding what actually mattered.

OpenAlex

  • Docs: openalex.org
  • Developer / API: api.openalex.org/works
  • SDKs & repos: Fully open catalogue of works, authors, venues and institutions; successor to Microsoft Academic Graph
  • Downloadable / offline: Entirely free API with no key required; complete database snapshots downloadable
  • Note: The only large-scale bibliographic database that is open all the way down, including bulk snapshots.

DBLP

  • Docs: dblp.org
  • Developer / API: dblp.org/faq/13501473.html
  • SDKs & repos: Authoritative CS bibliography — complete author and venue listings
  • Downloadable / offline: Free; full XML dump downloadable
  • Note: The fastest way to find everything a given researcher has published, and to see a conference’s full programme by year.

OpenReview

  • Docs: openreview.net
  • Developer / API: docs.openreview.net
  • Source: github.com/openreview
  • SDKs & repos: ICLR, NeurIPS, COLM and dozens of other venues — papers plus the full review threads
  • Downloadable / offline: Free; REST API
  • Note: Reading the reviews and author rebuttals teaches you how the field evaluates work. Nothing else exposes this.

CORE

  • Docs: core.ac.uk
  • Developer / API: core.ac.uk/services/api
  • SDKs & repos: Aggregates 300M+ open-access papers from repositories worldwide
  • Downloadable / offline: Free API and bulk datasets
  • Note: Good for finding the green open-access copy when a publisher’s version is paywalled.

Unpaywall

  • Docs: unpaywall.org
  • Developer / API: api.unpaywall.org
  • SDKs & repos: Finds legal free copies of paywalled papers via DOI
  • Downloadable / offline: Free API; browser extension
  • Note: Legal, author-deposited copies only. Install the extension and most paywalls simply stop appearing.

Papers We Love

  • Docs: paperswelove.org
  • Source: github.com/papers-we-love/papers-we-love
  • SDKs & repos: A curated, categorised repository of classic CS papers with local meetup talks
  • Downloadable / offline: Repo cloneable; many PDFs mirrored in-repo
  • Note: The best starting point if you do not yet know which papers matter in a subfield.

The Morning Paper (archive)

  • Docs: blog.acolyer.org
  • SDKs & repos: Adrian Colyer’s daily paper summaries, 2014-2021
  • Downloadable / offline: Free archive, still online
  • Note: No longer updated, but the back catalogue of ~1000 summarised papers is one of the great free CS resources.

USENIX Proceedings

  • Docs: usenix.org/publications/proceedings
  • SDKs & repos: OSDI, SOSP (co-published), NSDI, ATC, FAST, Security — the core systems venues
  • Downloadable / offline: Every paper free, immediately, with no membership. Often with recorded talks
  • Note: USENIX made everything open access years before the rest of the field. If a systems paper exists, check here first.

ACM Digital Library

  • Docs: dl.acm.org
  • SDKs & repos: SIGMOD, ASPLOS, PLDI, POPL, SoCC and the ACM journals
  • Downloadable / offline: Partly open: ACM Open and author-paid OA papers are free; others are paywalled
  • Note: 403s to automated clients; loads in a browser. For paywalled items, check arXiv, the author’s homepage or Unpaywall first — the free copy usually exists.

IEEE Xplore

  • Docs: ieeexplore.ieee.org
  • SDKs & repos: ISCA, MICRO, HPCA and the IEEE journals
  • Downloadable / offline: Mostly paywalled; abstracts free
  • Note: Almost always worth searching the author’s page or arXiv instead. Architecture authors in particular post preprints widely.

DROPS / LIPIcs (Dagstuhl)

  • Docs: drops.dagstuhl.de
  • SDKs & repos: ECOOP, ITP, CONCUR, SAT and many theory venues
  • Downloadable / offline: 100% open access, free PDFs, Creative Commons licensed
  • Note: A fully open publisher. Every paper, always free, no exceptions.

arXiv cs.AI / cs.MA

arXiv cs.SE (Software Engineering)

  • Docs: arxiv.org/list/cs.SE/recent
  • SDKs & repos: Coding agents, program repair, SWE-bench-style evaluation
  • Downloadable / offline: Free
  • Note: Where the coding-agent literature actually lives — easy to miss if you only watch cs.AI.

ACL Anthology

COLM

  • Docs: colmweb.org
  • SDKs & repos: Language-modelling venue carrying a growing share of agent work
  • Downloadable / offline: Open via OpenReview

OpenReview

  • Docs: openreview.net
  • Developer / API: docs.openreview.net
  • SDKs & repos: Reviews and rebuttals for agent papers
  • Downloadable / offline: Free; API
  • Note: Agent benchmarks attract methodological criticism; the reviews are where you find it.

SWE-bench

  • Docs: swebench.com
  • Source: github.com/SWE-bench/SWE-bench
  • SDKs & repos: The dominant coding-agent benchmark, with a public leaderboard
  • Downloadable / offline: Free; harness in the repo
  • Note: Read the harness before believing a score. Verified, Lite and Full are different tasks and get conflated constantly.

Terminal-Bench

  • Docs: tbench.ai
  • SDKs & repos: Benchmark for agents operating a real terminal
  • Downloadable / offline: Free
  • Note: Harder and less saturated than SWE-bench.

Berkeley Function-Calling Leaderboard

  • Docs: gorilla.cs.berkeley.edu/leaderboard.html
  • SDKs & repos: Standardised evaluation of tool- and function-calling ability
  • Downloadable / offline: Free
  • Note: The most useful single number for ‘can this model call tools reliably’.

Papers with Code

  • Docs: paperswithcode.com
  • SDKs & repos: Leaderboards and linked implementations
  • Downloadable / offline: Free

Hugging Face Papers

  • Docs: huggingface.co/papers
  • SDKs & repos: Daily curation, usually with the agent repos linked
  • Downloadable / offline: Free

Anthropic Engineering

  • Docs: anthropic.com/engineering
  • SDKs & repos: Practitioner write-ups including ‘Building effective agents’
  • Downloadable / offline: Free
  • Note: Currently the best-written practical material on agent design from any lab.

Google DeepMind publications

Microsoft Research publications

USENIX Proceedings

  • Docs: usenix.org/publications/proceedings
  • SDKs & repos: Where the systems and security consequences of agents are starting to appear
  • Downloadable / offline: All papers free
  • Note: As agents get deployment surface, the serious failure analysis will land here rather than at ML venues.

Education & reference implementations

Two tracks: Basic builds the foundations, Advanced is about reading and extending real implementations. Everything listed is free and publicly accessible.

Basic

21 resources across 6 topics.

Read this before you pick a framework

  • Building Effective Agents (Anthropic) — The most useful thing written on this topic. Its central argument — that most problems need a workflow, not an agent, and that you should start with the simplest thing that works — will save you months.
  • ReAct paper — The reasoning-and-acting loop every framework is a variation on. Short.
  • smolagents — About a thousand lines. Read it end to end and the magic disappears — an agent is a loop, a tool schema, and a stopping condition.

Learn the protocol layer

  • Model Context Protocol — The emerging standard for tool and data access. Learn it before any framework’s bespoke tool abstraction.
  • MCP specification — Short and readable. The trust-boundary section is the part most integrations get wrong.
  • MCP reference servers — Read two or three before writing your own.
  • OpenAI function calling — The raw tool-call interface underneath every framework. Most agent bugs are schema bugs.

A first framework

  • OpenAI Agents SDK — Small, clear, with tracing built in. Hides little, which is what you want while learning.
  • Pydantic AI — Type-safe and idiomatic Python. Good if you want validation from day one.
  • LangGraph — Move here when control flow and durability start to matter. The graph model is worth the learning curve.

Watch mature agents work

  • Claude Code docs — A production coding agent you can run and inspect. The hooks and permissions design is the lesson.
  • OpenHands — Fully open and capable. Read the sandboxed runtime.
  • Aider — The repository-map approach to context selection, clearly documented.
  • SWE-agent — The agent-computer-interface idea: tool design matters more than prompt wording.

Safety, from the start

  • OWASP Top 10 for LLM Applications — Excessive agency and insecure output handling are the entries that will bite you.
  • The lethal trifecta — Private data, untrusted content, external communication. Any two is fine; all three is not.
  • E2B — If your agent runs code it generated, sandbox it. This is a hard requirement, not a nice-to-have.

Finding and reading papers

  • Hugging Face Papers — Daily curation, usually with the agent repo linked alongside.
  • arXiv cs.SE — Where coding-agent research actually appears. Easy to miss if you only watch cs.AI.
  • SWE-bench — Read the harness, not just the leaderboard.
  • Anthropic Engineering — Practitioner write-ups with more signal than most preprints.

Advanced

33 resources across 6 topics.

Architecture & control flow

  • LangGraph — Explicit state graphs, checkpointing, interrupts, time travel. Read the checkpointer implementations.
  • Temporal — Durable execution done properly. Agent loops are long-running workflows; most frameworks reimplement this badly.
  • AutoGen — Event-driven multi-agent core, with research behind the design decisions.
  • Google ADK — Multi-agent composition plus a deployment story.
  • A2A protocol — Agent-to-agent delegation as a wire protocol rather than a library abstraction.
  • Building Effective Agents — Re-read it after you have shipped one. The workflow-versus-agent distinction lands differently then.

Context engineering

  • Aider repository map — Tree-sitter-driven repo mapping to fit a large codebase into a small context.
  • Claude Agent SDK — Compaction, subagents and context editing in a production harness.
  • SWE-agent — The ACI argument: design the tools for the model, not for the human.
  • LlamaIndex Workflows — Retrieval tightly coupled to agent control flow.
  • tree-sitter — The parsing layer under most code-aware context selection. See the compilers index.

Execution isolation

  • E2B — Firecracker microVMs per sandbox. Read the architecture, not just the SDK.
  • gVisor — Userspace kernel syscall interception — a different isolation point in the design space.
  • Firecracker — The microVM monitor underneath much of this. See the cloud index.
  • OpenHands runtime — A complete open implementation of sandboxed agent execution.
  • Modal — Fast cold starts matter when every tool call provisions a container.

Evaluating agents honestly

  • Inspect AI — Agent-aware solvers and scorers. The framework to use if results must withstand scrutiny.
  • SWE-bench harness — Read the evaluation harness. Contamination and variant confusion make most quoted numbers unreliable.
  • Terminal-Bench — Less saturated, harder to game.
  • Berkeley Function-Calling Leaderboard — Isolates tool-calling ability from everything else.
  • Langfuse — Trace every step. Agent failures are almost never visible in the final output alone.
  • OpenTelemetry GenAI semconv — Standard span names, so your traces survive a framework change.

Agent security in depth

  • MITRE ATLAS — Structure your threat model on an existing taxonomy.
  • OWASP GenAI — The deeper guides beyond the Top 10 list, including agentic-specific material.
  • garak — Probe your deployed agent, including tool-abuse paths.
  • PyRIT — Automated red-teaming with orchestrators and converters.
  • NeMo Guardrails — Action rails — constraining what an agent may do, not just what it may say.
  • Simon Willison’s archive — Keep reading it. This problem is not solved and the attack surface keeps growing.

Primary papers

  • ReAct — Reasoning interleaved with acting. The foundation.
  • Reflexion — Verbal self-critique as a learning signal across attempts.
  • Toolformer — Models learning when to call tools, self-supervised.
  • arXiv cs.MA — Multi-agent systems, where the coordination literature lives.
  • OpenReview — Agent benchmarks attract methodological criticism; the reviews are where you find it.

Video courses, channels & talks

9 resources across 1 group. Every channel and playlist below was fetched and title-verified on 2026-10-09. Handles drift and several plausible-looking handles resolve to the wrong channel, so a 200 response is not proof of identity — the links here were each checked against the channel title.

Channels & conference recordings

  • LangChain — Framework walkthroughs for agents and RAG; vendor content, verify against the docs.
  • Hugging Face — The open LLM ecosystem: transformers, datasets, agents, tool calling.
  • Dave Ebbelaar — Practical LLM and agent app builds, end to end, in Python.
  • Machine Learning Street Talk — Long-form research interviews; useful for context around current papers.
  • OpenAI — Model releases and DevDay talks; primary for API and product behaviour.
  • Google DeepMind — Research overviews and model announcements from DeepMind.
  • InfoQ — Conference keynotes and architecture talks — good for orientation, verify specifics elsewhere.
  • Latent Space — Agent-tooling interviews; current, but the frameworks discussed age within months.
  • TWIML AI Podcast — Applied-agent interviews, including RAG and evaluation practice.

Note: Framework channels dominate agent video and they age badly; watch for the pattern, then re-implement against current docs.

Conference videos, notes & archives

5 resources across 2 groups. Conference recordings are the primary-source tier of video: the speaker is usually an author of the paper, and where a talk exists the proceedings entry is often open at the same link. Every URL here returned 200 on 2026-10-09 unless the note says otherwise.

Agent work has no dedicated peer-reviewed venue yet; it appears at the main ML conferences and is demonstrated at framework dev days.

Conference channels & video archives

  • ICLR — Most agent-paper activity; the open reviews are useful for seeing what reviewers actually object to.
  • NeurIPS proceedings — The other half of agent research, especially tool use and evaluation.
  • ICML — Proceedings portal.
  • AI Engineer World’s Fair — The closest thing to a dedicated conference for applied LLM and agent work; recordings published by the organisers, vendor-heavy.

Notes, proceedings & paper-adjacent archives

  • Framework dev days are release notes — LangChain, OpenAI and Hugging Face agent talks describe the API as it was on the day. Re-check the docs before building on them.

If you only do three things

  1. Read Anthropic’s “Building Effective Agents”. Its argument that most problems want a workflow rather than an agent is the most load-bearing idea in the field.
  2. Read smolagents end to end. A thousand lines. Afterwards no agent framework can mystify you.
  3. Learn MCP from the spec, not from a framework’s wrapper around it.

Honest notes

  • Most things called agents should be workflows. A fixed pipeline with one model call per step is more reliable, cheaper and easier to debug. Reach for autonomy only when the task genuinely cannot be decomposed ahead of time.
  • Tool design beats prompt design. SWE-agent’s central finding. If the agent is failing, fix the tools and their schemas before rewriting the system prompt.
  • Agents escalate prompt injection into remote code execution. The lethal trifecta — private data, untrusted content, external communication — is an architecture constraint. No prompt fixes it.
  • Sandbox generated code. Always. E2B, gVisor or Firecracker. There is no safe version of running model-written code on your own machine.
  • SWE-bench scores are routinely quoted without naming the variant. Verified, Lite and Full are different benchmarks. Ask which, and ask about contamination.
  • This documentation ages in months. Several frameworks here have had breaking rewrites within a year. Trust the protocols and the papers over any individual framework’s API.