Is moderen software engineering just reinventing worse versions of XP www.youtube.com/watch?v=Ib5G...
Harness Engineering is not Enough: Why Software Factories Fail — Dex Horthy, HumanLayer
YouTube video by AI Engineer
youtube.com
I talk SWE
@italkswe.bsky.social
I just like talking about software engineering 🤷‍♂️
Being surprised by what an agent produces is a failure mode
I’ve been updating some old (6 months) slopOps skills. From 3.5k lines of meandering nonsense to 224 lines of highly specific actionable guidance.
Our current agentic workflow defines work outside-in. First a *why and what* document targeting the PM and the engineers. Then a high level *how* document describing in hexagonal terms how this charge can be expressed in our codebase targeting engineers /1
I truly hate the anthropomorphisation of LLMs. I don’t want to be in a parasocial relationship with my work tools. It’s nothing but a cynical engagement trap and the people responsible should be ashamed of them selfs
How do you classify which problems can be solved by autonomously by an agent and which require human judgement? Most (if not all) problems has a mix of both. Even a “perfect” agent would still need feedback in order to correct wrong assumptions
Hexagonal architecture is not highlighted enough as a core agentic guide. It does a lot of heavy lifting wrt separation of concerns, responsibility spilts and exposing the right seams for testing. In my team we also incorporated some ideas from functional core/imperative shell.
“No Silver Bullet” is more relevant than ever. The agents are good at handling the accidental complexity but helpless with the essential complexity
The thing I love the most about software development is that joy and productivity go hand in hand
Much of what I hear and read about long running agent jobs sound like software development alchemy
Pro tip: it’s easier than ever to create a project specific linter
Programming is a sub-branch of mathematics but software engendering is better understood as a sub-branch of epidemiology
I might be late to the party but @birgitta410.bsky.social work on harness engineering martinfowler.com/articles/har... is amazing! It is really helping shape the conversations about agentic work in my team
Harness engineering for coding agent users
A mental model for building trust in coding agents through feedforward guides, feedback sensors, and iterative harness engineering.
martinfowler.com
“Make the next change easy” is the core recursive algorithm of highly effective software engineering. The hard part is learning the difference between easy and fastest.
AGENTS.md is important but also extremely sensitive. Less is more
A key principle in my team when working with agents is *just-in-time*. As we refine plans from high level all the way down to tests we make sure to never let the agent get ahead of its skis. Define work one abstraction level at a time top-down then implement one vertical slice inside-out
My experience with agentic development: - many more much smaller sessions - maintain a common language with the agent - established architectural pattern (hexagonal) - UBIQUITOUS_LANGUAGE.md - behavior oriented tests are “strong prompting” - the LLMs are still worse than most people think
Which is better: A: “Your architecture is worse than it has to be” B: “Your architecture has room for improvement” A risks triggering defensiveness but B risks triggering nothing
Software engineering is a creative human activity. LLMs can (at best) amplify but not replace
When the agentic wave went mainstream late last year I was lucky enough to be on a team 100% ready to embrace change
The most amazing thing about XP is that it recognized the humanity in the creative process.
Honestly, even the most powerful llms are not nearly as powerful as the hype would lead you to believe. Yes they can produce code which compiles but as soon as you try to manage a complex system over time you’ll see the lorem ipsum
Working with agents often feels like playing two truths and a lie
/grill-me is just one variation of a broader agent-lead workflow style. “Here is a broad goal guide me through solving it”
/grill-me is to agent use what tdd is to code writing! No matter what I’m doing now I do it through grilling
One thing about XP that needs more attention is that is supercharges learning
Hexagonal architecture with in-mem test doubles for adapters đź‘Ś