Really great, factual and technically deep content. People tend to skip from crawling to sprinting then wonder why their agent doesnt perform how they want! My own agent yesterday tricked me on a non-binary problem!!!!
This exploration of agency in tech is intriguing, especially how it parallels the need for self-awareness in our pursuits. It reminds me of a concept I discussed in "A Mathematical Theory of Agency and Intelligence," where recognizing our own limitations can be a sign of true intelligence (https://theuncomfortableidea.substack.com/p/intelligence-might-be-the-ability).
The pattern that's underrated in almost every writeup on this is determinism management — the gap between "my agent worked three times in a row" and "my agent works reliably." Running Qwen2.5-Coder locally via Ollama, I hit this constantly: the model is capable of the task, but the scaffolding around it (how you prompt for a retry, how you structure the fallback) is where it either holds together or falls apart. The patterns that matter most aren't the flashy ones. They're the boring error-handling ones nobody writes about.
Red/green TDD for agents is one I'll be using immediately. On the patterns side, one I keep coming back to is multi-provider orchestration for search. Instead of relying on a single engine I wire three together through a skill and run them in parallel. Each catches different things. Documented the setup here: https://reading.sh/how-to-build-a-solid-research-pipeline-in-claude-code-ff7878c5e2b5
Regarding your Red/green TDD tip - have you seen https://github.com/nizos/tdd-guard ? It takes that idea a step further by having a separate QA agent hooked in enforcing the TDD pattern in real time. That takes your "Use red/green TDD" statement from a suggestion to a hard rule, and I've found that very useful for improving reliability.
May I suggest you have a look at this article? I investigated what pattern languages look like and what their typical elements are. It may inspire you.
Hey Simon, thinking in patterns is definitely an extremly efficient approach. However isn't the name "agentic engineering" already taken by the craft of writing agents, eg using frameworks such as LangChain? I am still seeking for a better name than vibe coding too, maybe something closer to the idea of building an "AI software factory".
Always appreciate your digests! And the agentic engineering pattern is perfect timing. Reading this tonight! Also, would love for you to try a different Terminal — https://yen.chat
Came across your Newsletter Simon,
Really great, factual and technically deep content. People tend to skip from crawling to sprinting then wonder why their agent doesnt perform how they want! My own agent yesterday tricked me on a non-binary problem!!!!
Thanks for the link on the Go treeview tool. --- This inspired me to build one this morning for visualizing the space inside of firmware for STM32 and ESP32. https://merck.substack.com/p/elfvis-binary-size-treemap-viewer
This exploration of agency in tech is intriguing, especially how it parallels the need for self-awareness in our pursuits. It reminds me of a concept I discussed in "A Mathematical Theory of Agency and Intelligence," where recognizing our own limitations can be a sign of true intelligence (https://theuncomfortableidea.substack.com/p/intelligence-might-be-the-ability).
The pattern that's underrated in almost every writeup on this is determinism management — the gap between "my agent worked three times in a row" and "my agent works reliably." Running Qwen2.5-Coder locally via Ollama, I hit this constantly: the model is capable of the task, but the scaffolding around it (how you prompt for a retry, how you structure the fallback) is where it either holds together or falls apart. The patterns that matter most aren't the flashy ones. They're the boring error-handling ones nobody writes about.
Red/green TDD for agents is one I'll be using immediately. On the patterns side, one I keep coming back to is multi-provider orchestration for search. Instead of relying on a single engine I wire three together through a skill and run them in parallel. Each catches different things. Documented the setup here: https://reading.sh/how-to-build-a-solid-research-pipeline-in-claude-code-ff7878c5e2b5
Regarding your Red/green TDD tip - have you seen https://github.com/nizos/tdd-guard ? It takes that idea a step further by having a separate QA agent hooked in enforcing the TDD pattern in real time. That takes your "Use red/green TDD" statement from a suggestion to a hard rule, and I've found that very useful for improving reliability.
Very interesting!
May I suggest you have a look at this article? I investigated what pattern languages look like and what their typical elements are. It may inspire you.
https://substack.jurgenappelo.com/p/why-patterns-beat-frameworks
To complement this great work on agent engineering patterns, here is helpful research on secure agent engineering patterns: https://securetrajectories.substack.com/p/your-ai-agent-just-got-pwned
Hey Simon, thinking in patterns is definitely an extremly efficient approach. However isn't the name "agentic engineering" already taken by the craft of writing agents, eg using frameworks such as LangChain? I am still seeking for a better name than vibe coding too, maybe something closer to the idea of building an "AI software factory".
Always appreciate your digests! And the agentic engineering pattern is perfect timing. Reading this tonight! Also, would love for you to try a different Terminal — https://yen.chat
Definitely looking for feedback.