September 2026

Things I’m seeing/reading/watching/listening to in September 2026…

Videos

• Shared clip (Facebook)
  • Nobody should have to motivate you…
• AI Skills with Matt Pocock (The Pragmatic Engineer Podcast)
  • Matt Pocock on why the “grill-me” skill has taken off and why software fundamentals matter more than ever in the age of coding agents.

Podcasts

Posts

• Git at any scale
  • Cursor on the challenge of hosting Git repositories at scale, introducing Continuity - a Git storage system that uses write-ahead logs in S3 for fully consistent, horizontally scalable repos, learning from the strengths and limits of earlier designs like GitHub’s Spokes.
• Harness Engineering for AI Agents in 2026
  • Vishal Mysore on the discipline of harness engineering - Agent = Model + Harness - and the programmatic wrapper (information, control and governance layers) that closes the gap between agent demos and production, where the vast majority of prototypes otherwise die.
• The AI-Native SDLC Playbook
  • Anthropic’s guide to integrating Claude across all six SDLC stages, from planning to maintenance - as the build phase accelerates, the bottleneck moves “to the steps to the left and right of the build phase,” so governance and specification matter more.
• Code Is No Longer the Bottleneck. Requirements Are.
  • Simon Martinelli critiques Anthropic’s AI-Native SDLC Playbook - agreeing code isn’t the bottleneck, but arguing it oversimplifies requirements into one agent session, and advocating the AI Unified Process with proper use cases, entity models and explicit non-functional requirements.
• Fleet Leadership
  • On the emerging role of “Fleet Leads” - senior engineers who orchestrate teams of AI agents and hold the system understanding machines can’t retain, as human judgment concentrates in context management, reviewing critical decisions, and architectural knowledge.
• Making Your Data Ready for Agentic AI
  • Pramod Sadalage and Prem Chandrasekaran argue agents need different data architectures than human analysts - humans work around bad data, agents act confidently on whatever they get - laying out four layers: trusted data via contracts, traceability, semantic context, and controlled operational access.
• Extreme Programming 1999 -> 2026
  • Anton Zaides revisits Kent Beck’s 1999 “Extreme Programming Explained” to see which of XP’s values, principles and practices still hold up in an era of AI and remote work - and what it still says about morale, quality and respect.
• How LLMs Actually Work
  • A comprehensive walkthrough of transformer-based LLM internals - tokenization, embeddings, attention and feed-forward networks - on the premise that “understanding the transformer machinery gets you most of the way there.”
• The Cost YAGNI Was Never About
  • Kent Beck reframes YAGNI: it was never about saving typing effort. It’s about preserving the optionality to build the right structure once requirements are clear, and not pulling costs forward while delaying revenue - which is why it still holds even when code generation is cheap.
• GraphRAG for Codebases: When do we actually need AI?
  • George Jayaratnam proposes pairing LLMs with two deterministic knowledge graphs - one for database schemas, one built from code parsers over syntax trees - to feed models precisely targeted context while keeping structural validation outside the AI.
• Use Cases vs User Stories - Same Content, Different Outcome
  • Simon Martinelli argues that use cases and user stories may describe the same domain but produce very different outcomes: use cases centralise rules and show end-to-end flows explicitly, while user stories scatter behaviour across items with implicit flows that get harder to follow as systems grow.
• Practices I Abandoned with Agents: An Ode to Test-Driven Development
  • After 25 years of TDD, Adam Tornhill explains why he dropped it going AI-native - TDD’s tiny increments suited human cognition, but agentic work happens at larger feature-level iterations where the agent handles the implementation detail.
• AI/works 2.0 - August updates
  • Thoughtworks ships version 2.0 of its AI/works agentic development platform, adding stronger enterprise governance, security and observability, plus better legacy modernisation, code evaluation, and grounding in trusted organisational knowledge.
• Resist The Slop - a short guide on how not to be a miserable dev in the age of AI
  • François (Fleur) Levasseur offers three strategies for keeping quality and sanity when working with coding agents: automated guardrails and strict testing, detailed instructions treating the agent like a junior dev, and tight feedback loops that keep you in the loop rather than rubber-stamping big batches.
• Introducing System One Models & Jev
  • TypeSafe AI introduces Jev, its first “System One Model” - a model built for fast, structured decision-making inside software rather than chat, claimed to hit similar intelligence on System One tasks while being two orders of magnitude faster and more efficient.
• I write 100% of my code with agents now
  • Cole Medin on having moved to writing effectively all of his code through agents, and the workflow that makes that reliable.
• The most helpful career advice I ever got
  • David Kline shares the piece of career advice that stuck with him most.
• Two months after I left LangChain
  • A reflection posted two months after leaving LangChain, on what has changed since.

Books

• Writing Use Cases for AI (Leanpub)
  • Simon Martinelli’s book on writing use cases that work for stakeholders, engineers and code-generating AI agents alike - because “implementation is now cheap but expressing intent precisely is the bottleneck.”
• Master Software Architecture: A Pragmatic Guide (Leanpub)
  • Maciej “MJ” Jedrzejewski distils 13 years of experience into an eight-step, pragmatic path for architects - from fundamentals and domain discovery through deployment, testing, security and evolving systems over time, with real-world examples and exercises.

Other