Build without limits - SDLC Briefing, #6
Hi! I am Ricardo Sueiras, leading SDLC transformation at Indicium AI. That means I get to work with customers and help them understand how to use agentic AI to build software. New tools, practices and opinions turn up e
Hi! I am Ricardo Sueiras, leading SDLC transformation at Indicium AI. That means I get to work with customers and help them understand how to use agentic AI to build software. New tools, practices and opinions turn up every week, and Build without Limits cuts through the noise. In each edition we put together a short, curated briefing on what's worth reading, watching and worth trying. You can skim it in five minutes and dive deeper wherever something catches your eye.
Top of Mind
Top of mind this week has been integrating quality into the harness. Two posts came up that are worth reading that cover some of the things I am seeing. First up, We handed testing to AI. Here is what stayed human which continues Roman's journey on how to unlock review and flow bottlenecks in agentic engineering. There are a lot of great observations within this post, and I like the conclusions he draws at the end "The bottleneck keeps moving. This time, the way through it is not more testers, but better systems around the ones we have.". On a similar vein, Vincent shares his ideas on How to Stop a Coding Agent From Weakening or Faking Tests to Make Them Pass and I have tried some of these techniques so am fully supportive. Essential reading to give yourself more confidence in the tests your agents run.
Must reads this week
These were my favorite posts this week.
Which Claude model should you build with? Fable, Opus, Sonnet, or Haiku helps answer one of the most common things we help engineering teams with, how to think about models and the tasks they are doing.
Agent Harnesses. Thinking Harder Beat All of Them. a post that suggests setting a models thinking level in Claude Code, produces better outcomes that the various elaborate configurations that we create. This one raised my eyebrows more than a few times.
System Optimizations for Cost-Efficient Agent Harnesses a research paper that I came across this week that took a look at how to optimise your harness.It shares five techniques to optimise cost in agent executions. First, we find reading existing context is the largest source of agent cost in real-world executions even with caching, taking up to 73% of the total cost. Second, tool calls incur significant cost without increasing the token count, causing up to a 26Γ cost increase at the same token count. Third, tool call outputs are often information sparse, and the agent uses only 27% of tools' outputs on average. Fourth, agents often invest many turns in filesystem exploration, and up to 89% of agent turns can be spent searching the filesystem in retrieval-heavy tasks. Finally, agent traces contain a large number of parallel subtasks, and we find on average 1.3 to 2.7 independent subtasks that can run concurrently to reduce context sharing across subtasks. Read the paper to dive into each of these.
WebMCP, explained with a todo app if you have not yet dipped your toe into WebMCP, then why not take a look at this post and be inspired to try it out. WebMCP provides agents with a robust framework in how to interact with your web applications, that borrows from Model Context Protocol (MCP).
Watch this
Bartosz Ocytko, Executive Principal Engineer at Zalando, shares three years of lessons of how they were able to scale agentic engineering across 2,000 developers and 150 engineering teams.
Tools to try out
Third-party tools links are provided as-is. They belong to their owners, follow their own licenses, and are used at your own risk.
aws-emulator is an official emulator of AWS services from Amazon, that joins Localstack and floci in allowing developers to run everything locally. This week my timeline has been full of developers experimenting what you can do. So far, the responses have been mostly good. One thing to bear in mind is that its not intended as a tool for testing security, so IAM support is limited.
ace is an experimental new tool from Github that explores what multiplayer agentic development might look like. Defo check the docs on this one (starthere). Still getting my head around this one, so if you are ahead of me, I would love to chat!
Not a tool/repo, but Apple now lets you run XCode inside Claude Code which means you can do things like build, run and test your app, render SwiftUI previews, drive simulators and real iPhones, read build logs and errors, or search Apple's developer docs. You will need XCode on your Mac (latest version) and then use this.
Install in Claude Code: /plugin install xcode@claude-plugins-official
I enjoyed reading and trying out some of the stuff in Apple's Xcode Plugin for Claude Code: What It Does
From the field
This week I have been exploring the relationship and impact of developer experience on organisations that are adopting agentic AI.
Conway showed organisations the impact of structure on its ability to build applications. Over the past ten years and the move the Cloud and cloud native architectures, this caused many organisations to rethink their structure (Amazon) or implement new systems (reverse Conway manoeuvre). Accelerate (and many studies since) showed that organisations that had invested in developer experience, were the organisations able to ride the wave of innovation and value generation that Cloud promised.
In the age of agentic AI, Conway has been dialled up to 11, with pressure points surfacing up across the software development lifecycle. Agentic AI tools have the potential to improve the developer experience like no other previous technology. Agentic engineering however, depends and leans heavily on strong foundations. Those foundations (and investments) that organisations should have been building during the Cloud era.
This surfaces up at customers in many different manifestations. It might be pressure points surfacing up as one area moves faster and others fail to keep up, or it is equally where they are looking to apply AI to fix problems they already have across their SDLC. When zooming out, we see the same challenges that lead back to poor or underinvested developer experience being at the root of the problem.
Organisations that have invested in a strong developer experience are going into the AI SDLC era well placed to take advantage of the benefits agentic AI brings. Its an area that organisations should start to reflect where they are, and then plan investments. This is beginning to happen, and I am seeing a growth in developer experience teams being formed across business right now.
Developer Productivity for Humans: Navigating the Tensions of AI in the Software Development Lifecycle is well worth reading to find out more about this.
If you enjoyed this, then remember to subscribe to the IndiciumAI blog so that you can keep up to date with future editions of this newsletter, as well as other great content that technologists at IndiciumAI will publish.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β full credit and traffic to the original publisher.