Skip to content

Resources

Abstract

Curation, not accumulation. Every item below has a reason to trust it stated next to it, and every link was checked on the date shown. The avoid list at the bottom is as useful as the rest — most of it is material that ranks well and is quietly out of date.

Verified as of 2026-08-21.

How to judge a source yourself

The field produces more confident writing than it does reliable writing. Three filters that work:

Check the date, then check what changed after it. MCP had a breaking spec revision in July 2026. Anything on MCP written before that teaches an API that no longer exists, however good it was.

No disclosed methodology and no named author with shown work → discard. This removes most "Top 10 agent frameworks" content in one step.

Prefer sources that publish their own negative results. A vendor blog that says "we tried this and it did not work" is worth ten that only report wins.

Essential

If you read only these, you are in good shape.

Source Why trust it
Anthropic engineering blog The highest-density teaching corpus in the field, and the rare vendor blog where the substance outweighs the marketing. Start with Effective context engineering (Sep 2025), Writing tools for agents (Sep 2025), Effective harnesses for long-running agents (Nov 2025), Demystifying evals (Jan 2026).
MCP specification The protocol itself. Check the revision date every time — 2026-07-28 is a breaking redesign. The architecture page is better than the spec for first contact.
Building effective agents Still the best vocabulary for agent patterns. Read the banner — Anthropic marks it superseded. Terminology, not current practice.
Context Rot 18 models, six task families. The empirical basis for treating your context window as a ceiling rather than a budget. Vendor-published; pair with NoLiMa (ICML 2025) where it carries weight.
LLM Evals FAQ Husain & Shankar, continuously maintained. The most current practitioner document on evaluation, with actual numbers.
OWASP Top 10 for Agentic Applications The only vendor-neutral security taxonomy for agents, each item backed by a documented incident.
Simon Willison Highest-trust individual source in the field. No affiliate incentive, publishes corrections, dates everything. His lethal trifecta tag is the public incident registry.

The pair worth reading together

Published one day apart, from two companies shipping real agents, arguing opposite conclusions:

Reading them as a pair teaches the actual decision criterion — context sharing versus parallelisable read-heavy work — better than either alone. Cognition revised their position in 2026: what works is multiple agents contributing intelligence while writes stay single-threaded.

Read directly, not summarised

Source Note
ReAct The origin of interleaved reasoning and acting. Short, current, not superseded.
Agent Skills spec A genuinely cross-vendor format — one SKILL.md folder runs across Anthropic, OpenAI, Google, Microsoft and AWS tooling.
A2A protocol Linux Foundation, v1.0 April 2026. Worth knowing; see the scepticism below before building on it.
The Attacker Moves Second OpenAI, Anthropic and DeepMind jointly breaking 12 published injection defences at >90%. The three parties with most reason to claim otherwise.
Adding Error Bars to Evals Miller. Standard errors, paired differences, power analysis. Short and directly implementable.
AI Agents That Matter Princeton. Why accuracy-only benchmarking produced needlessly expensive agents.

Free courses that are actually current

Course Cost Note
Hugging Face AI Agents Course Free, certificate Deliberately multi-framework rather than vendor-captured. Actively maintained. Its Unit 2 certification Space has been unreliable — use the lesson content.
DeepLearning.AI — Agentic AI Free to audit The flagship structured course on multi-step tool-using systems.
DeepLearning.AI — Evaluating AI Agents Free to audit Evals are the bottleneck skill; this is the best structured treatment.
Microsoft Agent Framework docs Free For the .NET half of the world. GA'd 1.0 April 2026; successor to Semantic Kernel and AutoGen.

Code worth reading

Repo Why
mini-swe-agent The agent loop in ~180 readable lines, with cost and step limits made explicit. Runs against Ollama. The single best read before touching any framework.
claude-cookbooks patterns/agents/ is the workflow taxonomy as runnable notebooks. CI-enforced against both code rot and link rot.
12-Factor Agents Framework-agnostic production principles. A 2025 document — the principles held up, the tooling references did not.

Tools that run locally, free

Tool Note
Ollama The substrate for everything in this path.
Phoenix pip install arize-phoenix && phoenix serve — tracing and evals, no account. Elastic 2.0, not OSI open source.
promptfoo YAML evals in CI, local judges supported. Set the grader explicitly or it reaches for a hosted model.
MCP Inspector Debug an MCP server without wiring up a host.
pgvector The realistic production answer for vector search.

Avoid, and why

Not bad-faith — mostly good material that time moved past. It still ranks well.

Item Problem
Any MCP tutorial predating mid-2026 The 2026-07-28 revision removed the handshake and sessions, deprecated roots/sampling/logging, and dropped standalone SSE. The SDK renamed FastMCP to MCPServer. Most published MCP content teaches a dead API.
anthropics/courses 22k stars, no commit since Nov 2025, examples target retired Claude 3 model IDs. Contains no MCP and no agents content.
RAGAS as a dependency Last release Jan 2026, no commits since Feb, 563 open issues. The metric vocabulary is worth learning; the library is not worth depending on.
AutoGen Maintenance mode; last release Sep 2025. Microsoft Agent Framework is the successor. Still recommended in current listicles.
OpenAI Swarm Superseded by the Agents SDK, by its own README.
awesome-mcp-servers 90k+ stars and 3,000+ unmerged pull requests. Nobody is curating it. Use the MCP Registry.
langchain-ai/rag-from-scratch Heavily recommended, 14 months stale, pre-dates LangChain v1.
"Top 10 agent frameworks for 2026" listicles Confirmed pattern of 2024 SEO refreshes still recommending AutoGen and Swarm.
Medium / Towards Data Science as a brand The publication carries no trust signal. An individual piece by a named practitioner with a real repo can be fine; the masthead means nothing.
LinkedIn AI content Over 40% of long-form posts flag as fully AI-generated. Default-distrust the channel.
Any "we solved prompt injection" claim Contradicts OpenAI, Anthropic and Google DeepMind simultaneously. See Safety.

Worth knowing, worth scepticism

A2A is a well-designed spec solving a real problem — cross-organisational agent delegation with identity and long-running tasks. After twelve months it has 150+ supporting organisations, hyperscaler platform support, and no publicly named production deployments. Learn the Agent Card and task-lifecycle concepts; do not treat it as something you need.

GraphRAG buys corpus-level sensemaking that no chunk retriever can do at any k. It does not buy better factoid QA — two independent evaluations found vanilla RAG matching or beating it on single-hop retrieval, with RAG-plus-reranking winning decisively, at roughly 57x the indexing cost. Decide on query mix, not on corpus size.