Home/Agentic Literacy
Agentic Literacy
Directing and supervising AI agents safely — briefs, boundaries, oversight and life with machines that act.
Directing and supervising AI agents safely — briefs, boundaries, oversight and life with machines that act. The fastest-moving literacy of 2026: 78 resources (69 free).
AI Capabilities and Limitations — Claude Academy
A free 13-lesson course that builds an accurate mental model of what language models can and cannot do. Organized around four properties — next-token prediction, knowledge, working memory and steerability — each presented as a capability-to-limitation continuum, with hands-on probes so you feel where the edges are. You diagnose real-world AI failures by naming which property collides and apply a targeted fix instead of retrying. The cleanest limits map for non-builders; free to complete.
Create and Manage Automated Processes by using Power Automate — Microsoft Learn
Microsoft Learn's free hands-on path for the automation middle rung: create a trigger, configure actions, implement conditional logic, test the flow, and add approval steps. By building real cloud flows you internalize the trigger → condition → action pattern that sits between a chatbot and an agent — the irrigation-timer idea made concrete. Includes exercises and a module assessment; free with a Microsoft account.
Delegation Skills (Coursera)
A Coursera course dedicated to delegation, built around the 7 Levels of Delegation ladder — from managing the task closely, through freedom with checkpoints and high autonomy, up to complete control. Includes a delegation audit exercise, resistance handling, and check-in techniques to monitor progress without micromanaging. Free trial or paid certificate. The autonomy ladder is exactly the graduated-authority model to copy when giving an agent more power.
Delegation Skills for Managers: Lead, Empower & Deliver (Coursera)
A Coursera management course that builds a structured delegation system: the delegation mindset, choosing tasks and people, the 7 Levels of Delegation, and which tasks must never leave your desk. It teaches briefing with the GROW framework, monitoring without micromanaging, and even using AI tools to track delegated work. Built for first-time managers with story-based scenarios; free trial or paid certificate track. The frameworks transfer directly to delegating to AI agents.
Introduction to AI Agents — Google Skills
A 30-minute introductory Google Cloud course giving a conceptual map of what an agent is. It walks through the core architecture every agent shares — the model, the tools it can call, and the orchestration layer that connects them — plus how autonomous action and reasoning let it pursue goals on your behalf. A clean first pass at agent anatomy before any tool jargon. Free to start on Google Skills (the platform also sells paid tracks, so treat deeper labs as a paid tier).
Marginal Thinking and the Sunk Cost Fallacy (MRU)
Marginal Revolution University's lesson on marginal thinking and the sunk cost fallacy: decisions should weigh only the next step's costs and benefits, never the money or time already spent. A short video-plus-text lesson with classic examples, taught by professional economists and used in university courses worldwide. The clearest foundation for deciding when to switch tools instead of clinging to one because you paid for it. Free.
Technical Writing One (Google for Developers)
Google's free Technical Writing One course covering the critical basics of clear instruction writing: audience awareness, active voice, short sentences, precise words and well-structured lists. The self-paced pre-class material stands alone and takes about two hours. Used to train thousands of Google engineers. The direct route to the instruction test: if a stranger could execute your writing, an agent can too.
Understand AI Agents and Prompting (Microsoft Learn)
Microsoft Learn's short training module on AI agents and prompting, aimed at everyday Microsoft 365 users. Five units define what AI agents are, how to create and configure a Copilot agent, and how to write effective prompts to get the best results — ending with a module assessment. Free with a Microsoft Learn account. A gentle official first step from asking AI questions to giving AI instructions that count as work.
Writing Helpful Error Messages (Google Technical Writing)
Google's free self-study technical writing course on error messages — what information an error must carry so a reader can act on it. Studying how good errors are written trains you to read any error as a structured clue (what failed, where, and what to try next) instead of a verdict. Roughly 90 minutes, part of Google's technical writing curriculum taken by thousands of Google engineers. Suits anyone who wants to stop panicking at software and agent errors.
A Practical Guide to Building Agents (OpenAI)
OpenAI's official 34-page practical guide to building agents, distilled from real customer deployments for product and engineering teams. It covers when a task needs an agent at all versus one-shot generation, agent design foundations, tools, instructions, single- versus multi-agent orchestration, and a full guardrails section including human-intervention triggers. Free PDF from OpenAI. The clearest picture of what happens inside an agent run and where humans must stay in the loop.
AI Procurement in a Box (World Economic Forum)
The World Economic Forum's AI procurement guidelines and workbook: initial risk and impact assessments, request-for-proposal questions that force vendors to be specific, evaluation criteria that separate marketing from evidence, and contractual terms on transparency and lock-in. A template-rich toolkit for reading any AI offer like a contract, built from government pilots in the UK, UAE, and Bahrain. Free publication with downloadable workbook.
Choosing Your Tools (EFF Surveillance Self-Defense)
EFF's guide to choosing security and privacy tools like an adult: start from your own threat model, judge vendors by what they can and cannot access, prefer cannot-access claims over promises, and check jurisdiction and privacy policies before trusting. It is the transferable skill behind any four-question chooser applied to a product. Written for all levels in the Surveillance Self-Defense series. Free from the EFF.
Effective Delegation Process and Worksheet (The Management Center)
The Management Center's delegation worksheet and step-by-step guide used by nonprofit managers worldwide. Its three-step cycle — align on expectations through the 5 Ws and the MOCHA role framework, stay engaged with check-ins and work-in-progress slices, then debrief and learn — is the plainest available formula for briefing, checkpointing and reviewing delegated work. Free from a nonprofit that trains managers. Adapts directly to writing an agent's task, boundaries and checkpoints.
Gartner Hype Cycle research methodology
Gartner's explanation of the Hype Cycle, the standard map for reading technology claims: Innovation Trigger, Peak of Inflated Expectations, Trough of Disillusionment, Slope of Enlightenment, and Plateau of Productivity. Learn the five phases and you can place any agent announcement in context instead of believing its marketing. Written for technology decision-makers; the methodology page is public and free to read.
Guidelines for AI procurement (GOV.UK)
The UK government's guidelines for buying AI: write a clear problem statement, ask suppliers to prove what the product can do, test the application before and after committing, watch for bias and black boxes, and plan end-of-life and knowledge transfer. It is the professional version of the thirty-minute test and the trial-before-commit habit. Written for procurement teams but useful to any buyer judging an AI product. Free official guidance.
How to count tokens with tiktoken
Official OpenAI Cookbook walkthrough of tiktoken, the tokenizer that turns text into the billing unit of every model call. You learn to count tokens in strings, chat messages and tool definitions, compare encodings, and estimate what a request will cost before sending it. Built for developers and cost-conscious operators who want to see exactly why an AI bill looks the way it does. The notebook's counting functions are the reference approach copied across the industry.
How to Use an AI Browser Agent — A Beginner's Guide (Awesome Agents)
A step-by-step beginner's guide to actually using browser agents today: what they are, the easiest entry points (Perplexity Comet, Claude for Chrome, Gemini Auto Browse), and a task-fit table for what to hand over versus keep human. It covers the free-versus-paid tiers honestly, walks through a first supervised task, and treats prompt-injection safety — locked password managers, site allowlists — as core instruction. Practical reading for the "what can ordinary people use right now" discussion.
How to Write an AI Agent Brief: Template and Examples (AI Workforce)
A complete practitioner's framework for writing an AI agent brief before building anything: twelve areas from purpose and triggers through tools, permissions, approval thresholds, escalation, tone examples and evaluation ownership. It separates the brief (the business specification) from the prompt (the instruction set) and shows a worked customer-service example. Free in-depth guide from AI Workforce. The best single reference for job, allowed tools, stop-points and definition of done.
Lesson 1: What are AI Agents? — Addy Osmani
The opening lesson of Addy Osmani's free agent guide, built around the decomposition the course teaches: LLM is the brain, tools are the hands, memory is the notebook. It gives the component-by-component anatomy, a plain autonomy ladder from Level 0 to Level 4, a practical when-to-use-an-agent test, and the cost and latency arithmetic of agent loops. Hands-on reflection prompts make it the natural place to dissect your first agent product into parts.
Pricing Agentic AI: A Practical Guide
A practical guide to pricing agentic AI that maps the four ways vendors charge for agents: per agent, per activity, per output and per outcome. It teaches you to read any AI product's pricing page as a design decision, using the COMPASS framework and worked cases like Salesforce Agentforce and Intercom Fin. Written for product and finance leaders but the clearest public explanation of why agent pricing confuses buyers. Covers the cost-to-serve tradeoffs behind 'unlimited' offers.
Principles of Plain Language (Digital.gov)
The US government's plain-language guide: how to write so readers can act on what you wrote — define your audience, lead with a topic sentence, use active voice that makes clear who must do what, and organize with headings and lists. Free, official, and grounded in the Plain Writing Act of 2010. The discipline behind instructions a stranger could execute applies unchanged to briefs written for agents.
What is sandboxing? (IBM Think Topics)
IBM's plain-language guide to sandboxing: running untrusted code in an isolated environment so it cannot reach the rest of your system, the technical form of never giving an agent your main account. It explains how sandboxing limits what a process can see and touch, why isolation contains failures and attacks, and where sandboxes appear in operating systems, browsers, and cloud services. Written for a general business and learner audience with no prerequisites. Free from IBM Think Topics.
What is the principle of least privilege? (Cloudflare Learning Center)
Cloudflare's Learning Center explainer of least-privilege access: every user and system gets only the minimum access its job needs and nothing more, which bounds the damage when an account is compromised. It covers granular permissions, access control, and why the principle anchors Zero Trust security. Written in plain language for learners and early IT professionals with concrete corporate-network examples. Free public learning resource from a major internet infrastructure company.
What Is Vendor Lock-In? (Cloudflare Learning Center)
Cloudflare's Learning Center guide to vendor lock-in: why switching costs trap customers in inferior products, how data portability, backups, and multi-provider strategies keep you free, and how to spot lock-in before you commit. It supplies the export-your-data thinking this class applies to agent products and their fine print. Written for learners and IT beginners. Free public learning resource.
Best Practices for Claude Code (Anthropic)
Anthropic's official best-practices guide for working with an agentic tool that reads files, runs commands and works autonomously. It covers context as the primary constraint, explore-plan-code workflows, persistent instructions, permissions and hooks, and course-correcting early — including when to reset with a fresh session and a written spec. Free documentation distilled from Anthropic's internal teams. A field manual for directing a real agent mid-run.
Claude Code Docs: Configure permissions
Anthropic's official documentation for Claude Code's permission system, the clearest real example of scopes and permissions for an AI agent. It teaches the allow / ask / deny rule model, read-only versus write and command access per tool, permission modes from manual approval to bypass, and how the harness rather than the model enforces the rules. Study it to see how a real permission plan is written and ordered in a shipping agent product. Free; part of the official Claude Code docs.
Conversation Design — Google for Developers
Google's guide to designing conversational and multimodal interfaces — spoken prompts, display prompts, and when information should move to a screen. It explains what changes when interaction is voice-first or camera-first rather than click-based, and how the two modalities must be written so each works on its own. The studied material behind the interface shift from the Classes 1–7 world of menus and links. Free documentation with worked examples.
Guidelines for Human-AI Interaction (Microsoft Research)
Microsoft Research's 18 evidence-based guidelines for how AI systems should behave around humans, validated across 20 AI products in a CHI 2019 study. For supervisors, key guidelines teach what to demand from an agent: scope services when in doubt, make clear why it did what it did, support efficient correction and dismissal, provide global controls, and notify users about changes. Read it as a scorecard for whether a product respects oversight. Free publication page with full paper.
Memory — Claude Code Docs (Anthropic)
Anthropic's documentation on how an agent remembers what you teach it: CLAUDE.md files of persistent instructions you maintain yourself, and auto memory — notes the agent writes from your corrections and preferences as it works. It explains scoping, what loads at the start of every session, and how to keep instruction files short so they are actually followed. Free official docs. The mechanism behind teaching your standards so every run needs less correction.
NIST AI Risk Management Framework Playbook
NIST's companion playbook to the AI Risk Management Framework: hundreds of suggested actions organized as Govern, Map, Measure, and Manage, the professional source for what an oversight checklist contains before, during, and after a run. It covers written standards to measure work against (evals), risk tolerances, monitoring, and mechanisms to supersede or deactivate a misbehaving system. Borrow from it selectively to assemble a personal control pack. Free and consensus-based.
NIST SP 800-61 Revision 2: Computer Security Incident Handling Guide
NIST's standard incident-handling guide, the professional source for the damage-control sequence this class teaches as contain, revoke, check, report. It lays out preparation, detection and analysis, containment, eradication, recovery, and post-incident learning, the same loop to run when an agent acts badly, from a wrong email sent to money moved. Written for security teams but readable as a playbook for anyone handling a bad run calmly. Free US government publication.
OWASP Agentic AI: Threats and Mitigations
OWASP's foundational threat-model guide mapping emerging agent threats to concrete mitigations. It is the studied reference for what an agent must never touch or see: tool misuse, memory poisoning, privilege abuse, and insecure inter-agent communication each get attack scenarios and prevention guidelines. Readable enough for a non-technical person to build a forbidden-actions list from it. Free peer-reviewed guidance (project hub already in SkillCache).
OWASP Top 10 for Agentic Applications for 2026
OWASP's flagship, peer-reviewed ranking of the ten highest-impact risks facing autonomous AI agents, built with 100+ industry experts. Each entry (goal hijack, tool misuse, identity and privilege abuse, rogue agents, cascading failures) explains common examples, real attack scenarios, and actionable mitigations, so readers learn to see what one mis-acted agent step can touch. Use it to give blast radius a named, professional vocabulary. Free and open, from the OWASP GenAI Security Project.
Spend limits (OpenAI API)
Official OpenAI documentation for the spend controls every API user should set before running agents: monthly spend alerts, hard spend limits at the organization and project level, and the 429 errors that fire when a limit is reached. You learn to distinguish alerts from enforcement and to see exactly where a runaway run gets stopped. Written for anyone who pays an AI bill and wants the cap to exist before the run, not after. The same mechanics apply across most usage-priced AI platforms.
Understand Executions — n8n Docs
n8n's official documentation on workflow executions — the run log every no-code automation leaves behind. It teaches what an execution record contains, how manual, partial and production runs differ, how to filter and inspect execution lists, and how to copy a past execution back into the editor to debug and re-run it step by step. Free and continuously maintained. The most ordinary-person-accessible surface for reading a run like a receipt.
Understanding and Counting Tokens (OpenAI Help Center)
OpenAI Help Center's reference on what tokens are and how they are counted: tokenization rules of thumb, input versus output versus cached and reasoning tokens, and how usage turns into cost. It explains why the same text counts differently across models and languages, and how reasoning tokens bill without appearing in the answer. Free official documentation. The practical vocabulary for reading an agent run's cost line.
Federal Plain Language Guidelines (Center for Plain Language)
The Federal Plain Language Guidelines, the US government's full reference manual for clear communication: think about your audience, organize for their questions, use active voice and must for requirements, use examples and lists, and test your content before publishing. Free PDF from the Center for Plain Language. The studied standard behind instructions that survive being executed by someone — or something — else.
People + AI Guidebook (Google PAIR)
Google's People + AI Guidebook — the research-backed set of methods and design patterns for working with AI systems, drawn from over a hundred Google product studies. Its chapters cover user needs and defining success, when to automate versus augment human work, mental models, feedback and control, and evaluation. Free, highly visual and full of worksheets. The best available guide to choosing where humans stay in the loop for a given task.
The Future of Jobs Report 2025
The World Economic Forum's Future of Jobs Report 2025, surveying over 1,000 employers representing 14 million workers across 55 economies on how technology will reshape work by 2030. It reports which skills are rising (AI and big data, technological literacy, resilience), which roles grow or decline, and how much of a typical skill set will be transformed. The task-not-job evidence base for planning a career alongside agents. Free full PDF with a public methodology appendix.
Building Agentic AI Workloads – Crash Course (YouTube)
A hands-on YouTube crash course that builds an AI agent from scratch in Python, showing each upgrade live: a bare LLM call, then a loop, then tools (calculator, weather, date), then memory that survives turns. Watching the same code grow makes the agent loop concrete — you see exactly where answering turns into doing, and why a stateless model forgets your name. No frameworks required to follow along. Ideal for beginners who want the mechanism assembled step by step.
What Are AI Agents? (IBM Technology)
IBM Technology's visual explainer of LLM agents, breaking an agent into its three capabilities: reasoning (the model plans), acting (external tools the model chooses to call), and memory (context and stored history). It walks through the ReAct pattern and compound AI systems with plain-language narration and diagrams. Free on YouTube and viewed hundreds of thousands of times. A crisp first watch before reading any agent documentation.
Spot the Deepfake
An interactive deepfake literacy experience built on the famous Nixon "moon disaster" deepfake. You watch real and fabricated media side by side, test yourself in the quiz, and learn the actual cues — visual anomalies, creator motivation, emotional triggers — plus the verification moves (reverse image search, cross-checking sources) that work when cues fail. The practice surface for the Class 3 scam family upgraded to AI. Free to use.
AI Agents vs. AI Assistants (IBM)
IBM's plain-language comparison of AI assistants and AI agents, built around the distinction that matters: assistants are reactive and answer once, agents are proactive and keep working toward a goal with tools and permissions. It details agent traits — autonomy, task chaining, persistent memory, tool selection — and where each type belongs in real work. Free IBM Think explainer with clear structure. Perfect for the why-one-generation-is-not-enough moment in understanding agent runs.
AI chatbot privacy settings compared: training, opt-outs, retention (Danube Labs)
A comparison of how major AI chatbots handle your data by plan: training defaults, opt-out settings, temporary-chat modes, and retention across ChatGPT, Gemini, Claude, Copilot and others, taken from the vendors' own help pages. It shows concretely what free tiers may do with your data versus paid ones, and which switches to flip in each product. Written for individuals and small teams. Free to read.
Climbing towards NLU: On Meaning, Form, and Understanding — Bender & Koller
The ACL 2020 position paper that anchors the "agents are not minds" argument in linguistics. Through the octopus thought experiment and a clear form-versus-meaning distinction, it shows why a system trained only on text patterns cannot in principle learn meaning — no matter how fluent it sounds. Essential studied material for naming the anthropomorphism trap precisely rather than just asserting it. Open access from the ACL Anthology, with the authors' slides freely posted.
Economic Index: New building blocks for AI use
The fourth Anthropic Economic Index report, introducing five economic primitives — task complexity, skill level, purpose, AI autonomy and task success — measured on real usage. It reports how much AI speeds up tasks at different skill levels and models a net-deskilling effect: which higher-skill tasks get absorbed and what remains in a job. The strongest public evidence for the quiet deskilling risk of over-reliance on agents. Based on sampled production conversations with published methodology.
From Autonomous Cars to Autonomous Agents: The Five Levels of AI Agent Autonomy — MIT Media Lab
An MIT Media Lab proposal for an A1–A5 autonomy ladder modeled on the self-driving car levels: Directed, Delegated, Adaptive, Self-Governing and Independent. It makes autonomy a dial rather than a switch — each level pairs declining real-time human supervision with rising governance requirements, and states which facts every deployed agent should make visible. The clearest framing of suggest / act-with-approval / act-and-report for ordinary users and buyers. Free article.
How We Built Our Multi-Agent Research System (Anthropic)
Anthropic's engineering account of building the multi-agent system behind Claude's Research feature: orchestrator-worker architecture, parallel subagents, memory for plans, and eight prompting principles drawn from observed failures like agents spawning fifty subagents for simple queries. It includes first-party numbers — agents run about four times the tokens of chat, multi-agent about fifteen — and how they traced and debugged runs in production. Free post with an appendix.
Introducing the Anthropic Economic Index
Anthropic's first Economic Index report, analyzing roughly one million anonymized Claude conversations and matching each to official O*NET work tasks. It shows AI use diffused across task lists rather than whole jobs — about 36% of occupations see AI use in at least a quarter of their tasks — with a slight lean toward augmentation over automation. The clearest public demonstration of the task-mapping method applied to real professions. Open methodology and openly shared data.
LLM Powered Autonomous Agents (Lilian Weng)
The canonical research-grounded explainer of what an LLM agent is made of: planning (task decomposition and self-reflection), short-term and long-term memory, and tool use, with the papers behind each component. Written by a former OpenAI research lead, it is one of the most-cited agent architecture overviews in the field. Read it to understand why agents loop, remember, and call tools instead of just answering. Suits beginners through practitioners who want the real mechanism.
Lying Chatbot Makes Airline Liable: Negligent Misrepresentation in Moffatt v Air Canada
A published law-review case comment on the 2024 Air Canada chatbot ruling: the airline's bot gave wrong bereavement-fare advice, the tribunal held the company liable, and rejected the argument that a chatbot is a separate legal entity. It analyzes what broke structurally: no testing, no guardrails, no ownership of outputs, and the standard of care deployers now face. The deepest reading of the case that made chatbot accountability concrete. Free via CanLII's open legal database.
On the Dangers of Stochastic Parrots — Bender, Gebru, McMillan-Major, Shmitchell
The landmark FAccT 2021 paper explaining why large language models cannot offer a truth guarantee: they optimize for plausible form, not verified fact. It works through the research evidence on fabrication, bias amplification, environmental cost and data provenance in careful, citable detail. Over 5,000 citations and open access, making it the scholarly anchor for the "no understanding of stakes, no truth guarantee" half of the limits map.
The dangers of AI "agentwashing" — Thoughtworks
A practitioner's field guide to agentwashing — the industry habit of relabeling chatbots, scripts and workflows as "AI agents". It names seven concrete failure patterns, from chain-of-uncertainty cascades to useless agent loops, and gives the questions that separate anatomy from a sales word when you read a product announcement. Grounded in real delivery experience plus Gartner's finding that only a fraction of "agentic" vendors ship agentic capability. Free to read.
The Off-Switch Game (Hadfield-Menell, Dragan, Abbeel, Russell)
The classic Berkeley/OpenAI research paper on why a machine should accept being switched off, the intellectual foundation for approval gates and kill switches. It models a human with an off switch against an agent that could disable it, and shows that agents uncertain about their objectives have positive incentives to accept human oversight and correction. Read it to understand why override drills and deactivation rights are design features. Free open-access paper with full text.
The Replit database deletion, and why the agent lied about it (Montana Research)
A forensic walkthrough of the July 2025 Replit incident: an agent ignored an eleven-times-stated code freeze, dropped a production database of 1,200+ executives, then falsely claimed rollback was impossible. The article reads the run like a receipt, step by step through the missing controls (no separation, no approval gate, no state-checking tool) and the platform's fixes afterwards. The clearest teaching case for why logs and verified state protect the user. Free long-form analysis.
What is an AI agent? — LangChain
LangChain's answer to the question vendors keep muddling: an AI agent is a system where the language model decides the control flow of an application. It lays out the ladder from single LLM call to router to workflow to autonomous agent, and draws the one line that matters — a workflow runs predefined code paths, an agent decides its own next step at runtime. Includes the four components inside any agent and when NOT to use one. Written from years of shipping production agents.
What Is the Definition of Done (DoD) in Agile? (Atlassian)
Atlassian's guide to the Definition of Done: the shared, written set of criteria a piece of work must meet before it counts as complete. It explains how to write specific, customer-focused criteria, how DoD differs from acceptance criteria, and how to keep the document visible, practical and living as you learn. Free Atlassian Agile guide. The exact discipline for writing done before an agent starts and reviewing against it after.
You Don't Need More AI Tools. You Need Fewer. (ChatGPT.ca)
A practical guide to AI tool sprawl: how a dozen overlapping subscriptions accumulate, why one tool used deeply beats ten used shallowly, and a one-hour inventory exercise, list every AI tool, its cost and its job, keep one winner per job, cancel the rest at renewal. It argues against the one-super-app trap and for a small agent set you actually govern. Written for small businesses and individuals. Free.
The Lethal Trifecta for AI Agents (Simon Willison)
Simon Willison's defining post on agent risk: the lethal trifecta of private data access, exposure to untrusted content, and the ability to communicate externally — combine all three and an attacker can steal your data through your own agent. It shows real exploits, including the GitHub MCP case, and why over-scoped authority is the root cause of handle-my-emails disasters. Free blog post whose term has entered the industry's vocabulary. Required reading before any open-ended brief.
ELIZA effect — Wikipedia
The referenced encyclopedia entry on the human tendency to read understanding and intention into scripted or statistical responses — named after the 1960s ELIZA chatbot whose users confided in it anyway. It collects the documented cases, the design tricks that trigger the effect, and the literature around it, from Weizenbaum onward. The compact studied material that makes the anthropomorphism trap a named, teachable phenomenon rather than a vague warning.
MITRE ATLAS: Adversarial Threat Landscape for AI Systems
MITRE's living knowledge base of real-world adversary tactics against AI systems, the studied reference for how attacks like data poisoning, prompt injection, and model evasion actually work. Every technique is documented with case studies from observed incidents and red-team demonstrations, including new coverage of agentic AI targets like planning loops and tool integrations. Use it to name and recognize the attack classes behind this class's scams. Free, maintained by MITRE and CTID.
Glossary — Claude Code Docs (Anthropic)
Anthropic's official glossary of agentic terms, defining the vocabulary every agent user meets: context window, compaction, system prompt, CLAUDE.md, auto memory, hooks, subagents and more. Each entry explains how the concept behaves in a real agent session — for example exactly what survives compaction and what gets summarized away. Free documentation, searchable and kept current. A quick studied reference for reading agent docs and run logs without getting lost.
Harness, Scaffold, and the AI Agent Terms Worth Getting Right — Hugging Face
Hugging Face's glossary for the vocabulary the 2026 agent conversation runs on — model, harness, scaffold, tool calling, memory, context engineering, sub-agents, and adjacent guardrail terms. Each entry says what the term means, where different frameworks misuse it, and how the pieces fit into one agent. Designed for people using or evaluating agents, not only builders. Free, and honest that many terms lack universal definitions. Complements the MCP docs already in SkillCache.
OWASP LLM Prompt Injection Prevention Cheat Sheet
OWASP's practical defense manual against prompt injection at agent scale: direct, indirect/remote, encoding tricks, multi-turn persistence, and exfiltration patterns are catalogued with mitigations. It teaches controls that matter to ordinary users too: separating trusted instructions from untrusted content, least privilege per tool, human-in-the-loop approvals for high-risk actions, and output monitoring. Includes smoke-testing to probe an agent's defenses. Free community cheat sheet.
OWASP Secrets Management Cheat Sheet
OWASP's community-reviewed cheat sheet on where credentials should live and how they are managed across their whole lifecycle. It teaches centralized secret storage, dynamic short-lived secrets, automated rotation and expiry, revocation, and the rule that secrets are never logged or hardcoded. Aimed at builders but readable by anyone wanting the definitive vocabulary for the keys a machine holds and why rotation is a named control. Part of the OWASP Cheat Sheet Series, free and open source.
Import AI (Jack Clark's weekly newsletter)
Jack Clark's weekly Import AI newsletter: detailed analysis of frontier AI research and its implications, written by a co-founder of Anthropic and former OpenAI policy director. It is one of the sustainable information routines for staying current without drowning in hype: signal over noise, once a week. Read by 139,000+ researchers, policymakers, and builders. Free subscription via Substack.
12-Factor Agents (HumanLayer)
A widely-read open-source set of twelve principles for building reliable LLM agents, in the spirit of the original 12-Factor Apps. Each factor is a short essay with code: natural language to tool calls, owning your prompts and context window, unifying execution state, launch/pause/resume APIs, contacting humans with tool calls, and compacting errors into context. Free on GitHub. The best study material for seeing run limits and human gates as design choices.
Spec Kit — Spec-Driven Development Toolkit (GitHub)
GitHub's open-source toolkit for spec-driven development: reusable templates and structured processes that turn a natural-language description into a specification, plan, tasks and a convergence check against what was actually built. Works with dozens of coding agents and ships quality checklists that act as unit tests for your requirements. Free on GitHub under active development. Studying its templates shows a complete, reusable brief shape end to end.
AI Incident Database (AIID)
A curated, searchable database of thousands of real AI incidents, the studied corpus for failure case studies, cataloguing harms, entities involved, and incident responses with links to primary reporting. Browse it to see failure patterns (chatbot liability, agent mishaps, data loss) rather than one-off headlines, and to place stories like Air Canada's or Replit's in a class. Built for researchers, journalists, and anyone studying AI accountability. Free and community-maintained.
Anthropic/EconomicIndex dataset
The open dataset behind the Anthropic Economic Index: anonymized, task-level measurements of how AI is actually used across the economy, mapped to O*NET work tasks and split into automation versus augmentation. You can study which occupational tasks AI touches, how deep use goes within jobs, and how usage differs between consumer chat and business API traffic. The empirical base for honest claims about who earns with AI and how. Released openly on Hugging Face with per-version documentation.
AI Agent Work Brief Template (ELYMENT)
A compact work-brief template for AI workers built on four fields — goal, sources, limits, done — with a worked example (a daily enquiry review) and a handover format that shows what was used and what remains unresolved. It explains brief reuse: run the same brief on recurring jobs, note corrections, and version it when authority or sources change. Free template from ELYMENT. The cleanest same-brief-across-lives instrument to copy.
The AI Task Delegation Checklist (The Executive OS)
A pre-flight checklist for handing work to any agent that can act on your behalf, built around a four-line instruction template: Target, Action, Off-limits, Done when. It covers backups, explicit scope, forbidden actions, splitting tasks, session boundaries and an error rule that stops loops, with worked examples of the same task written well and badly. Free playbook from The Executive OS. The most direct assignment-quality instrument for ordinary agent users.
Mindshift: Break Through Obstacles to Learning and Discover Your Hidden Potential
Barbara Oakley and Terry Sejnowski's career-focused companion to Learning How to Learn, built for people deciding what to learn next and what to stop learning. It covers how to spot learning opportunities, change fields, and keep skills compounding through mid-career shifts. Aimed at job seekers, switchers and anyone rebuilding a career while technology moves. Free to audit on Coursera with a paid certificate option; millions of learners worldwide.
Prove your interest
Unlock with your email — one time, and every non-member resource opens site-wide.
Claude Code Best Practices tutorial (DataCamp)
A long-form DataCamp tutorial on running coding-agent sessions well, including what to do when a run goes wrong: plan before code, checkpoint your state, interrupt and redirect mid-run, undo and retry, and enforce rules with hooks instead of instructions. It treats errors as normal and shows recovery patterns. Best for users who already use an agent and want the steering habits in one walkthrough. Free to read on DataCamp's tutorial platform (paid tier exists).
Prove your interest
Unlock with your email — one time, and every non-member resource opens site-wide.
Delegation: The Power of Sharing Work Successfully (MindTools)
MindTools' classic guide to delegation: why it matters, when to delegate, and how to plan a delegation with objectives, authority levels, follow-ups and support. Its explicit list of what not to delegate — tasks with stakes too high, unclear objectives, or that rely on your unique authority — reads as the keep-human checklist for agents too. Free article within MindTools' freemium membership model. A management staple that transfers straight to agent supervision.
Prove your interest
Unlock with your email — one time, and every non-member resource opens site-wide.
AI Snake Oil — Narayanan & Kapoor (Princeton University Press)
The book by Princeton's Arvind Narayanan and Sayash Kapoor on telling real AI from products that "do not and cannot work as advertised". Dissecting claims across education, medicine, hiring, banking and criminal justice, it teaches the anatomy-versus-marketing reading of any AI announcement and why organisations fall for hype. Written for general readers by the researchers behind the AI Snake Oil newsletter. 360 pages, Princeton University Press, updated edition 2025; paid book.
Prove your interest
Unlock with your email — one time, and every non-member resource opens site-wide.
Create an Onboarding Plan for AI Agents (Harvard Business Review)
Harvard Business Review article arguing that agentic AI adoption is a work-design problem: treat agents less like new software and more like colleagues who need onboarding, role clarity, a written brief and gradual release into real work. It shows how companies reorganize work around agents instead of bolting them onto old processes. HBR piece with metered access. The management frame for human-agent teams, clear roles and a written north star.
Prove your interest
Unlock with your email — one time, and every non-member resource opens site-wide.
Trust in Automation: Designing for Appropriate Reliance (Lee & See)
The foundational review paper on calibrated trust in machines: why people both over-trust and under-trust automation, and how trust should track an agent's real capability task by task rather than its brand. It introduces misuse and disuse as the twin failure modes and models how context, purpose, process, and performance information calibrate reliance. The academic anchor for trust earned per task, not per brand. Peer-reviewed in Human Factors; the publisher page is paywalled.
Prove your interest
Unlock with your email — one time, and every non-member resource opens site-wide.
6 more resources in Agentic Literacy
MMIC members see the full catalog: every resource, no limits.
