Skip to content

Knowledge base

What we tell teams to read next.

Other people's work on agentic engineering and the organizations around it, with our notes and our book ratings.

  • 296 sources
  • 18 with our notes
  • 140 books
Type

296 sources

Our recommendations

Each one first ran as a spotlight post on LinkedIn.

  • Channel youtube.com

    Leon van Zyl (opens in a new tab)

    Teams that stall while designing a whole orchestration up front can copy his order: ship the thin happy path first, then harden where the harness shows a real failure.

    Our full note

    Leon van Zyl teaches agent building step by step, in code and in no-code tools, moving between agent frameworks, Claude Code and Cursor. In his build-along tutorials he starts from an empty project, gets the simplest version running, watches where it breaks and fixes that. Teams that stall while designing a whole orchestration up front can copy that order: ship the thin happy path first, then harden where the harness shows a real failure.

    LinkedIn, 14 September 2026 (opens in a new tab)

  • Blog tidyfirst.substack.com

    Tidy First? (opens in a new tab)

    With agents producing code at speed, a team that never decides on purpose when to tidy ends up with a codebase that grows fast and changes slowly.

    Our full note

    Kent Beck's newsletter arrives in small, frequent pieces, from deep technical dives to short reflections on the trade. Its title asks whether to clean up messy code before a change, and Beck treats the answer as an economic decision: tidying costs a little now and pays back on every later change in the same spot. With agents producing code at speed, a team that never makes that call on purpose ends up with a codebase that grows fast and changes slowly.

    LinkedIn, 7 September 2026 (opens in a new tab)

  • Channel youtube.com

    DevOps & AI Toolkit (opens in a new tab)

    An agent is one more platform user, and every path that still needs a senior engineer to unblock it stops the agent as well.

    Our full note

    Viktor Farcic's channel works through real DevOps, Kubernetes and CI/CD products and talks openly about their tradeoffs, including when something is not worth it. His running thread on internal developer platforms rests on one mechanism: an expert codifies what they know into a service, so nobody has to wait for them again. An agent is one more platform user, and every path that still needs a senior engineer to unblock it stops the agent as well.

    LinkedIn, 31 August 2026 (opens in a new tab)

  • Blog maggieappleton.com

    Maggie Appleton (opens in a new tab)

    Context files for agents rarely say how settled they are, and her stages let an agent tell a validated rule from a guess written under deadline.

    Our full note

    Maggie Appleton, a designer and anthropologist, writes illustrated essays on programming, knowledge and how people think with what they build, kept as a digital garden she tends over years. Her essay on the garden's ethos labels each note by stage, from seedling to budding to evergreen, with the dates it was planted and last tended. Context files for agents rarely say how settled they are, and her stages let an agent tell a validated rule from a guess written under deadline.

    LinkedIn, 24 August 2026 (opens in a new tab)

  • Channel youtube.com

    IndyDevDan (opens in a new tab)

    Once skills, agents and prompts spread across repositories and people, sharing them becomes a distribution problem, and The Library answers it the way package managers do for code.

    Our full note

    A practitioner channel on building with coding agents, with one working pattern a week shown end to end and the reasoning left in. Start with The Library, a skill that keeps a team's other skills in one catalog of references, with commands to add, pull and sync them. Once skills, agents and prompts spread across repositories and people, sharing them becomes a distribution problem, and The Library answers it the way package managers do for code.

    LinkedIn, 17 August 2026 (opens in a new tab)

  • Book teamtopologies.com

    Team Topologies (opens in a new tab)

    Start with the Thinnest Viable Platform: the same question keeps a team's agent harness from becoming the bottleneck it was built to remove.

    Our full note

    Matthew Skelton and Manuel Pais give engineering leaders a vocabulary for arranging teams so value flows between them, in a book now in its second edition and in patterns, templates and assessments on their site. The idea to start with is the Thinnest Viable Platform: the smallest platform that removes real friction for the teams on it, grown only when a need pulls it. The harness and automation around a team's agents form a platform too, and the same question keeps them from becoming the bottleneck they were built to remove.

    LinkedIn, 10 August 2026 (opens in a new tab)

  • Book web.stanford.edu

    A Philosophy of Software Design (opens in a new tab)

    Its contrast of deep and shallow modules gives reviewers the words to ask for depth in code an agent wrote to pass its tests.

    Our full note

    John Ousterhout's short book grew out of a Stanford course and has one subject: complexity in software and how to fight it. The chapter to read contrasts deep modules, which hide a lot behind a small interface, with shallow ones whose interface is almost as complicated as what it wraps. An agent writes code that passes its tests without feeling that difference, so reviewers need these words to ask for depth and check for it.

    LinkedIn, 3 August 2026 (opens in a new tab)

  • Blog simonwillison.net

    Simon Willison (opens in a new tab)

    Count the lethal trifecta for every agent you run, and design out the full set before anything goes near production.

    Our full note

    Simon Willison has documented the practical edges of these models in public, post by post, with no vendor behind him. His lethal trifecta names the combination that makes prompt injection dangerous: an agent with access to private data, exposure to untrusted content and a way to send data out. Count those three for every agent you run, and design out the full set before anything goes near production.

    LinkedIn, 27 July 2026 (opens in a new tab)

  • Blog anthropic.com

    Building effective agents (opens in a new tab)

    Its five patterns and the line between paths you wrote and paths an agent chooses tell a team where to put a gate.

    Our full note

    Anthropic's essay collects what the company learned building agents with dozens of teams: the implementations that worked used simple, composable patterns instead of heavy frameworks. Its middle section names five patterns, prompt chaining, routing, parallelization, orchestrator-workers and evaluator-optimizer, each with a diagram and a note on when to use it. It also separates steps that follow paths you wrote from an agent choosing its own path, which tells a team where to put a gate.

    LinkedIn, 20 July 2026 (opens in a new tab)

  • Research getdx.com

    DX Research (opens in a new tab)

    Usage rose 65 percent, median PR throughput just under 8 percent, and DX traces most of the gap to work outside the editor.

    Our full note

    DX asked the same questions across more than 400 engineering organizations over 16 months while AI tools rolled in. Usage of those tools rose 65 percent, median PR throughput just under 8 percent, and these were mature adopters. DX traces most of the gap to work outside the editor, such as review and integration, and its writing on AI-native engineering organizations points at the fix.

    LinkedIn, 13 July 2026 (opens in a new tab)

  • Channel youtube.com

    Tessl (opens in a new tab)

    Where a team's intent lives only in tickets and chat threads, an agent rewriting a module works from a guess.

    Our full note

    Tessl builds spec-centric software, and its channel hosts the AI Native Dev podcast with Guy Podjarny and Simon Maple, plus conference talks and practitioner panels. The question under the episodes: once an agent can regenerate code from a clear description of intent, the specification becomes the thing a team protects. Where that intent lives only in tickets and chat threads, an agent rewriting a module works from a guess.

    LinkedIn, 6 July 2026 (opens in a new tab)

  • Tool github.com

    Spec Kit (opens in a new tab)

    Study it for the spec as a contract: when a lot is at stake, it settles what done means once, before the agent writes a line.

    Our full note

    GitHub's Spec Kit has a team write a spec, a plan and a task breakdown before the agent builds, and it works with most coding agents. It sits between prompting the agent directly, which wins for spikes, small fixes and exploration, and heavier frameworks such as BMAD. Study it for the spec as a contract: when the problem is understood and a lot is at stake, it settles what done means once, before the agent writes a line.

    LinkedIn, 29 June 2026 (opens in a new tab)

  • Channel youtube.com

    Ona (opens in a new tab)

    An agent that cannot install packages, run the tests or build the project cannot check its own work, however strong its model.

    Our full note

    Ona is an agent platform whose agents run in isolated cloud environments and open pull requests on their own, and its channel carries short demos, getting-started walkthroughs and recorded virtual events. Start with the Background Agents Virtual Summit, where the talks keep returning to one finding: the environment matters as much as the agent. An agent that cannot install packages, run the tests or build the project cannot check its own work, however strong its model.

    LinkedIn, 22 June 2026 (opens in a new tab)

  • Tool code.claude.com

    Hooks reference (opens in a new tab)

    The Stop hook puts a deterministic gate where a prompt could only ask.

    Our full note

    Claude Code's hooks reference documents shell commands the harness runs at fixed points in the agent loop, such as before a tool call or when the agent tries to finish. Read the Stop hook first: it can run the tests and send the agent back to fix failures before it may hand over. That puts a deterministic gate where a prompt could only ask, and the docs also show how a hook exiting the wrong way silently lets the agent stop.

    LinkedIn, 15 June 2026 (opens in a new tab)

  • Research softwareengineeringproductivity.stanford.edu

    Software Engineering Productivity Research Group (opens in a new tab)

    Read it before the next internal AI report, since a count of logins and acceptance rates says nothing about whether a team ships faster.

    Our full note

    This Stanford group studies what AI does to software engineering productivity, working from telemetry across many companies instead of surveys. Its method keeps usage, such as seats activated, prompts sent and lines accepted, apart from outcome: working software that reached a user. Read it before the next internal AI report, since a count of logins and acceptance rates says nothing about whether a team ships faster.

    LinkedIn, 8 June 2026 (opens in a new tab)

  • Blog jedi.be

    Patrick Debois (opens in a new tab)

    Teams now run code an agent wrote and few people read closely, and his writing on ownership, observability and fast feedback applies directly.

    Our full note

    Patrick Debois, who coined the term DevOps, now writes on his personal blog about the agentic version of the problem DevOps answered. Start with his essay on self-tuning context, where agents rewrite their own instructions and raise operations questions about review, drift and rollback. Teams now run code an agent wrote and few people read closely, and his writing on ownership, observability and fast feedback applies directly.

    LinkedIn, 1 June 2026 (opens in a new tab)

  • Blog dbreunig.com

    Drew Breunig (opens in a new tab)

    Four named failure modes let a team diagnose a misbehaving agent precisely and fix its context on purpose.

    Our full note

    Drew Breunig writes long, careful essays on how language models behave: the data, the context and the mechanics under the demo. His essay How Long Contexts Fail names four failure modes: poisoning, distraction, confusion and clash. With those names a team can diagnose a misbehaving agent precisely and fix its context on purpose.

    LinkedIn, 25 May 2026 (opens in a new tab)

  • Blog anthropic.com

    Anthropic Engineering (opens in a new tab)

    Every harness component encodes an assumption about what the model cannot yet do alone, which gives teams a reason to review and prune their own harness as models improve.

    Our full note

    Anthropic's engineering blog describes, in engineering detail, how a model lab builds the harness around its models. Its post on harness design for long-running work splits planning, building and evaluating across three agents, so the builder never grades its own output. The post also holds that every harness component encodes an assumption about what the model cannot yet do alone, which gives teams a reason to review and prune their own harness as models improve.

    LinkedIn, 18 May 2026 (opens in a new tab)

Books to start with

The top of our list for technology leaders, rated by relevance from 10 to 1.

All 140 books Curated Book List

By topic