Masters3d

Technical. Creative. Tactical. Director.

Software Engineering

How the work, the craft, and the career of software engineering change when agents do much of the execution.

17 posts

The Last 20% and the Next Nine

📅 October 15, 2026 ⏱️ 9 min read

The first 80% of a feature gets the demo, the congratulations, and the promotion packet. The last 20% (the part that makes the thing actually usable) costs a magnitude more and is rewarded far less. The reason is structural: every additional nine of reliability is a different engineering problem than the one before it, and you cannot reach it by extending a design that was never built for it. In the age of AI, when the token cost of writing the scalable version is roughly the token cost of writing the disposable one, the V1-then-rewrite habit deserves a second look.

Read more →

AI, SI, and the Problem With Renaming a Term of Art

📅 October 08, 2026 ⏱️ 9 min read

The top labs and the current administration got together and decided to start calling it Super Intelligence instead of Artificial Intelligence, SpaceX is renaming its AI unit to SpaceXSI, and some states like California are sticking with AI. This is about the name itself rather than the politics. I actually like the SI framing, the one real argument for it is linguistic because super and intelligence travel across languages and the acronym does not flip the way IA does in Spanish, I am wary of anything that shades into compelled speech, and my only real concern is whether the benefit reaches everybody rather than the top one percent.

Read more →

Models Are Tools: Portability, Watermarks, and the Work I Want to Keep

📅 September 26, 2026 ⏱️ 11 min read

A few weeks of running Astra next to Opus 5 left me with three conclusions: the behavior I depend on is mostly harness and habit rather than intelligence, watermarking every output treats a shared piece of work as if the model made it alone, and I want a personal profile that travels with me so the model stops being the thing I have to learn.

Read more →

There Is No Speedup for Shared Understanding

📅 September 23, 2026 ⏱️ 7 min read

Six months of building AI infrastructure and coaching coworkers on their agent workflows taught me the bottleneck is no longer code creation. Someone can write ten thousand lines in a day, but nobody found a way to make a colleague understand those ten thousand lines any faster than in 2019. Validation is still the vehicle through which humans actually learn a system, and that work has not been automated.

Read more →

The Why Doc and the How Doc: Steering Agents from the Middle

📅 July 31, 2026 ⏱️ 7 min read

Start with the code works for the software engineering circle of the Venn diagram, but a technical program manager naturally starts from the why, not the code. Once agents act like subordinates, the person steering them sits in the overlap and needs two documents: a Why doc facing outward toward the problem and stakeholders, and a How doc facing inward toward the agent's technical direction.

Read more →

Verification-Driven Development: Let Proofs of Concept Write the Spec

📅 July 29, 2026 ⏱️ 10 min read

Spec-driven development is a vague enough term that it can mean anything from a paragraph of intent to a fully waterfalled design. I tried the fully-specified version and it backfired: it left no room for an autonomous agent to explore alternatives. What worked instead was naming the actual sequence. Build multiple proofs of concept, verify each one against something real, and only then write the design document, informed by what the proofs of concept actually showed.

Read more →

The GUI Is Becoming an Output

📅 July 24, 2026 ⏱️ 7 min read

For thirty years we optimized software to be easy for a human hand. Agents invert that goal. The graphical interface is turning from the thing you drive into the thing you read, and the tools that survive the shift are the ones an agent can talk to.

Read more →

Shift Left Until the PR Is Just a Confirmation

📅 July 10, 2026 ⏱️ 11 min read

Agentic tooling made pull requests cheap and plentiful, so the bottleneck moved from writing code to trusting it. The answer is not another review loop but front-loaded rigor: invest the engineering effort ahead of time to build a full, production-shaped synthetic environment you run locally, so that validation itself becomes cheap and the PR turns into a context-sharing mechanism instead of an approval gate. Tools like .NET Aspire and Dapr show what that inner loop can look like, and the same principles carry over to homegrown systems.

Read more →

9 Years of Copious Notes

📅 June 29, 2026 ⏱️ 6 min read

Nine years of work notes, from sort-of-daily Markdown files in 2017 to monthly Word docs to a single yearly running document, and how AI agents finally took over the writing through worklogs.

Read more →

10,000 Hours Was Never Enough

📅 June 27, 2026 ⏱️ 8 min read

Time matters, but expertise does not come from time alone. In software engineering, AI makes the missing part impossible to ignore: growth comes from new constraints, real feedback, rising stakes, and ownership of outcomes.

Read more →

Language Choice in the LLM Era

📅 May 24, 2026 ⏱️ 10 min read

A practical guide to choosing programming languages for frontend, CLI, and server development with decision matrices that account for modern LLM-assisted development.

Read more →