Seb

The biggest cost driver for coding agents building features doesn't seem to be context size itself, but the number of model round-trips. Recent example: 3 agents (Architect, Implementer, Verifier) working on a small improvement to an existing feature in a 70k LOC codebase - would have taken me 5–8 hours manually. The agents took 3 hours, and none of their contexts exceeded 150k tokens. But the team required 1,200 model round-trips. Final cost: 260 million tokens. Does that make sense? 🤔

What was true before is even more important when agents do the coding: Consistent software architecture is more valuable than perfect architecture. 😉

My coding agent instructions say: - Follow KISS - Be minimalist I even added the supreme design principle: “Perfection is achieved, not when there is nothing more to add, but when there is nothing left to take away.” And still ... 🤦‍♂️

GenAI may not be truly creative. But it can point you to existing ideas you would never have thought of. And often, that is enough to build great solutions. 😉

Exactly 💯😎

Code review is still crucial for software quality. My code review agent almost always finds something. 🤷‍♂️😉

Gherkin seems to be the best way to make agent-written specs and tests reviewable. But coding agents seem less efficient at working with Gherkin than with other languages. Anyone found a significant improvement? Linter? Compiler? Editor? Intermediate language? 🤔

Which multi-agent framework do you use? 🤔 I used github.com/unclebob/swarm-fo… by @unclebobmartin very successfully for quite a while 👍 Now working on my own which I'll release soon 😉

With coding agents, I am now building features I would never have built before. The effort required to learn how to build them would have been so high that it simply wouldn't have been worth it. Now it is. 😉

We used to refactor code. Now we refactor Markdown. 😉 When agents maintain prompts and skills, they tend to keep adding more text instead of simplifying what is already there. Over time, this creates "technical debt". So I now regularly run review and cleanup sessions to keep my agent instructions sharp, simple, and focused.

Treating coding agents as tools is the wrong mental model. A better one: A coding agent is like a skilled developer who is new to your team and project every day. That changes how you work with them. You focus on building a harness that provides the knowledge, rules, and feedback they need. 😉

I have multiple projects where almost all the code is written by AI. How much of your code is written by AI meanwhile? 🤔

Are you actually concerned about security risks when running your coding agents in /yolo mode? 🤔

Why is everyone so obsessed with Opus 5.5 - as if Astra and Sol don't exist 🤔

This tendency of current models to overengineer is pretty annoying 😤 My agent team worked for 3 hours - I just requested a full rollback and complete rework🤦‍♂️

How do you handle context compaction in your coding agents? I switched to a simple rule: "Always compact before starting the next slice." Gives me more control over the context. And so far, it seems to save tokens too. 😉

Have you seen getartcraft.com/apps? Photoshop, Illustrator, Excel, ... rebuilt with AI coding agents: "Clean-room: implemented from public specs and observed behaviour only." A small team can now use existing software as an executable specification to (re)build complex applications. Impressive. 🤯

Which quality metrics do you use to monitor the design your coding agents produce?🤔

Very nice 👇😉 LGTM (Looks Good to Me) - Claude Opus 5.5 music video youtu.be/3TNpOD6bov8?si=ixt5… via @YouTube