The reading list before deploying AI agents

When teams ask us "what should I read before starting with AI agents," we end up sending the same list of resources. We're going to consolidate it here.

This isn't a list of Avery NXR docs. It's a reading list across the AI agent space — papers, posts, books, frameworks — that we think helps people think clearly before they deploy.

Everything here is something we've actually read and found useful. We're not curating to be impressive. We're curating to be useful.

Conceptual foundations

"The Bitter Lesson" by Rich Sutton (essay, 2019).

The single most-cited AI essay of the modern era. Argues that general methods leveraging computation beat specialized hand-engineered methods, over the long term.

Why read it: helps you understand why scale matters in AI, but also why specialized small models can still win for specific workloads in 2026.

"Anthropic's Building Effective AI Agents" (2024).

Anthropic's framework for thinking about agent design. Workflows vs agents. When to use which. Common patterns and antipatterns.

Why read it: the language and framework here have become industry-standard. Most agent discussions implicitly use this taxonomy.

"Reflexion: Language Agents with Verbal Reinforcement Learning" (paper, 2023).

One of the foundational papers on agent self-reflection patterns. Agents that learn from their mistakes.

Why read it: foundational for understanding why some agent designs perform better than others. Influences how Avery NXR's audit ledger + confidence scoring works.

"ReAct: Synergizing Reasoning and Acting in Language Models" (paper, 2022).

Older but still important. Introduces the reasoning-then-acting pattern that most agent frameworks build on.

Why read it: gives you the foundational pattern. Once you've read this, most agent framework docs make more sense.

Practical guides

"Building LLM applications for production" by Eugene Yan (post series).

Extensive practical writing on deploying LLM applications. Patterns, antipatterns, things that work, things that don't.

Why read it: more grounded than most "production AI" writing. Eugene has actually shipped this stuff.

"Tools and Agents" by Simon Willison (ongoing blog).

Simon writes prolifically about LLM and agent tooling. His blog is one of the best sources for practitioner-level thinking.

Why read it: he's curious, honest, and good at evaluating new tools. Reading his evaluations of LLM tools sharpens your own evaluation skills.

"Prompt Engineering Guide" (open source, prompt-engineering.org).

Comprehensive guide to prompt engineering patterns. Updated regularly.

Why read it: most agent failures we see come from prompt issues. This guide gives you the patterns to avoid common mistakes.

"How to evaluate AI agent outputs" by various authors.

There isn't one canonical reference here, but search around for evaluation frameworks. Hamel Husain has good practical writing on this.

Why read it: most teams skip evaluation and regret it. Knowing how to evaluate is what separates teams that get value from agents and teams that don't.

On local AI specifically

"Local LLM Resources" curated by Tim Dettmers (researcher).

Tim is one of the most respected voices in efficient LLM deployment. His curated resources are excellent.

Why read it: if you're considering local-first AI, his framing on quantization, hardware, model selection is foundational.

Ollama documentation.

If you're going to use Avery NXR or any local-first AI tool, you'll touch Ollama. Their docs are good.

Why read it: understanding Ollama removes mystery from local model deployment. 30 minutes of reading saves hours of confusion.

"Why Local AI" (various authors, no canonical source).

We've written about this extensively (see [posts 144, 149, 189]). Other thoughtful writing on the topic exists.

Why read it: helps you articulate to yourself (and stakeholders) why local-first matters and when it doesn't.

On AI agent failures

"Agent failures in production" by various practitioners.

There's no single canonical source. We wrote [post 179] on this. Look for first-hand accounts from practitioners.

Why read it: you'll hit failures. Reading about others' failures helps you anticipate yours.

"The AI alignment problem" (general topic).

You don't need to become an alignment researcher, but reading some baseline material helps you think about agent safety architecturally.

Why read it: helps you design safety into agents from the start, rather than retrofitting after a scary moment.

On agent ecosystem and category

"State of AI Agents" (annual reports by various analysts).

Multiple firms produce annual surveys of the AI agent space. Quality varies. Take with grains of salt.

Why read them: gives you a sense of how the broader market is moving. Helps with strategic thinking even if specific predictions vary.

Your competitors' blogs (yes, really).

If you're evaluating Avery NXR, read Lindy's blog, Relevance AI's blog, n8n's blog, CrewAI's docs. Different companies frame the agent space differently.

Why read them: triangulating across multiple framings helps you build your own mental model rather than adopting any one company's framing.

On AI economics and business strategy

"The State of AI" by Air Street Capital (annual report).

Long, comprehensive, free. Covers research, capabilities, businesses, and societal questions.

Why read it: helps you understand where AI is going beyond your specific tool choice.

"AI and the Cost Curve" by various economists.

Multiple writers have analyzed AI cost trajectories. Helps you reason about when costs change architectural decisions.

Why read it: cost trajectories drive a lot of architectural decisions. Understanding the curves helps you make better long-term bets.

Books worth reading

"Designing Machine Learning Systems" by Chip Huyen.

Not specifically about agents, but the foundational thinking applies. How to think about ML systems in production.

Why read it: if you're going to maintain agents long-term, you're operating a ML system. Chip's framework applies.

"AI Engineering" by Chip Huyen.

Newer book specifically on the LLM application era. Covers everything from prompt engineering to evaluation to deployment.

Why read it: most comprehensive practical book we've seen on this material.

"The Beginning of Infinity" by David Deutsch.

Not about AI. About knowledge, explanations, and progress.

Why read it: helps you think about what AI can and can't do at a foundational level. Useful for resisting hype in both directions.

What we'd skip

We won't name names, but there are categories of content we'd encourage you to skip:

→ Hype-cycle posts about "X kills Y" framings. Usually wrong. Always too simple.

→ Generic "future of AI" pieces by non-practitioners. Lots of these. Few are useful.

→ Vendor marketing dressed as thought leadership. Easy to spot once you've read enough.

→ AI-generated content about AI. The feedback loop is unhealthy. Stick to human-written analysis.

→ Twitter threads claiming AGI is imminent (or claiming it's impossible). Both extremes get attention. Neither is right.

How to actually use this list

We're not suggesting you read everything before deploying. That's a way to never deploy.

The right approach:

→ Read 2-3 foundational pieces before starting (Bitter Lesson, Building Effective Agents, ReAct) → Start building/deploying agents → Read more deeply as specific questions arise → Use the rest of the list as reference when you hit walls

The reading is in service of better deployment, not in place of it.

What we wish we'd known earlier

If we'd read this list before starting Avery NXR (we'd read most of it, but not all):

→ We'd have invested more in evaluation infrastructure earlier → We'd have prioritized audit transparency from day one (we did, but the conviction would have been deeper) → We'd have understood the local-first vs cloud-first trade-offs more clearly upfront → We'd have built better safety architecture in the first iteration

The reading list isn't a substitute for doing the work. It's how you do the work better.

The recommendation

If you're starting with AI agents:

→ Read 2-3 foundational pieces this week → Deploy something small (Avery NXR or similar) → Read more as questions emerge → Build with confidence in your foundations

If you've been deploying for a while:

→ Pick the 3-5 pieces above that you haven't read → Read them → Notice which ones change your thinking → Adjust your deployments accordingly

Either way, the reading is iterative. You won't be done. The space is too young for "done" to exist.

→ avery.software — Free Desktop tier. The reading list helps you deploy better. The deploying is where the actual learning happens.