AI agents & automation
AI reading list
An AI agent uses a model to choose actions, receives results from tools and continues toward a goal. The surrounding software decides what tools exist, what state survives interruptions and which actions need a check. Start with an explanation, inspect a small workflow, then ask whether the system actually completed the intended task.
People to read
Open the first material, then continue with the author on X. People appear once, grouped by their main reading use and ordered by handle within each group.
7 people
No people match that search. Try another name, subject or reading task.
Understand the agent architecture
Decide how a model, tools, planning and execution state fit together before choosing a framework.
-
Start by defining the agent’s environment, available tools and goal. Huyen’s book excerpt connects planning to actual function calls, then explains why tool choice, parameters, results and completion need separate checks. It is a useful primer before choosing a framework.
Start with this Web material · huyenchip.com
Read the overview, Tools and Planning sections, then the failure modes. The examples explain how the next action depends on tool results; they also show places to involve a person.
This work in the author’s profile
Published . Material checked .
-
Choose the surrounding software an agent needs. Chase separates framework abstractions from a runtime that preserves execution state and a harness with ready-made tools and instructions. His second essay helps decide when several workers are useful and how to coordinate their context.
Start with this Web material · langchain.com
Agent Frameworks, Runtimes, and Harnesses- oh my!
An accessible comparison of framework, runtime and harness responsibilities. The definitions are Chase’s proposed mental model, with overlapping boundaries; compare it with the needs of your own task.
This work in the author’s profile
Published . Material checked .
Connect tools to a workflow
Follow a tool call, a screenshot or a document through a working interface and into the next action.
-
Move from document Q&A to a process that updates other software. Liu traces a contract through parsing, policy retrieval, a structured report and an action connector. Typed handoffs and review points explain how to combine fixed business rules with flexible model decisions.
Start with this Web material · llamaindex.ai
Beyond chatbots: adopting Agentic Document Workflows for enterprises
Follow the four-stage table, then the contract-risk example. Notice the artifacts passed between stages and the separate places to review the clauses, matches, report and final action.
This work in the author’s profile
Published . Material checked .
-
Build a small tool-using loop you can inspect. Willison’s LLM tutorial exposes Python functions and plugins to a model, shows an incorrect SQL query followed by schema lookup, and prints tool calls. You can see the difference between requesting an action and the host program executing it.
Start with this Web material · simonwillison.net
Large Language Models can run tools in your terminal with LLM 0.26
Start with the built-in tool demo and --tools-debug output, then the read-only Datasette example. Python and terminal knowledge help; examples show the models and software version available in May 2025.
This work in the author’s profile
Published . Material checked .
-
See how an agent gets visual context from a desktop. Steinberger’s Peekaboo walkthrough explains a macOS screenshot tool with separate analysis and a CLI that can be called when needed. It offers a concrete interface-design example alongside his broader personal-agent work.
Start with this Web material · steipete.me
Peekaboo 2.0 – Free the CLI from its MCP shackles
Read How It Works and the capture/analyze examples. This is a macOS tool and a dated release walkthrough; screenshots and model-backed image analysis have different setup requirements.
This work in the author’s profile
Published . Material checked .
Check outcomes and control actions
Ask who may change state, who checks success and whether a completed task stays reliable across repeated runs.
-
Ask what an agent may change and who checks the result. Weng reviews persistent files, delegated workers and self-improvement loops, then explains why permission controls and evaluators should remain outside the parts the agent is optimizing. Her examples distinguish a plausible artifact from a completed task.
Start with this Web material · lilianweng.github.io
Harness Engineering for Self-Improvement
Start with Harness Design Patterns, then Future Challenges. The review’s discussion of reward hacking, independent checks and human decision points is useful even before the more technical optimization methods.
This work in the author’s profile
Published . Material checked .
-
Test whether an agent repeatedly completes the same request while talking to users and changing records. Yao’s coauthored τ-bench paper separates a correct database outcome from conversational variation and measures consistency across trials. It also states what outcome-only checks can miss.
Start with this Web material · arxiv.org
τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains
Read the airline example and section 3’s task, reward and pass^k definitions before the experiments. Yao is a coauthor. Simulated users and simplified policies make this a controlled research benchmark.
This work in the author’s profile
Published . Material checked . Authorship source.
A few bearings for your reading
These guides explain the people, language and questions you may meet along the way.
An AI agent demo checklist: tools, state and success
Assess an AI agent demo through its tools, state, access and checked outcome. Read a published database-error recovery and plan the next useful test.
Prompt injection in AI agents: follow the tool calls
Trace a published prompt-injection attack from a code comment to a private Slack read and public post, then inspect the permissions behind the tool calls.
How this list is selected
We selected people through their own writing, code and coauthored research. Each first read explains a concrete part of an agent system. Vendor authors describe tools they build; research benchmarks keep their original test conditions and team authorship. Recent X activity is recorded separately and was not assessed for this selection.