AI Enhanced Engineer

09/03/2026

𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗦𝗼𝗳𝘁𝘄𝗮𝗿𝗲: 𝗛𝗼𝘄 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁𝘀 𝗔𝗿𝗲 𝗥𝗲𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗶𝗻𝗴 𝘁𝗵𝗲 𝗦𝗼𝗳𝘁𝘄𝗮𝗿𝗲 𝗣𝗮𝗿𝗮𝗱𝗶𝗴𝗺

This new paper by Zhenfeng Cao presents the agent as the software itself and moves the human to intent architect. We read it closely. Five takeaways.

𝟭. 𝗖𝗼𝗱𝗲 𝗯𝗲𝗰𝗼𝗺𝗲𝘀 𝗮𝗻 𝗶𝗻𝘀𝘁𝗿𝘂𝗺𝗲𝗻𝘁, 𝗻𝗼𝘁 𝘁𝗵𝗲 𝘀𝘆𝘀𝘁𝗲𝗺. Traditional software encodes every decision before runtime. Agentic software generates decision logic on demand, writes code for the specific path, and discards it. The capability persists. The code does not.

𝟮. 𝗧𝗵𝗲 𝘀𝗰𝗮𝗹𝗶𝗻𝗴 𝗮𝗿𝗴𝘂𝗺𝗲𝗻𝘁 𝗶𝘀 𝗮𝗯𝗼𝘂𝘁 𝗰𝗲𝗶𝗹𝗶𝗻𝗴𝘀. Human cognitive capacity is fixed. Model capacity grows with compute. The paper frames this as the structural reason agents change which problems are economically tractable.

𝟯. 𝗧𝗵𝗲 𝗵𝘂𝗺𝗮𝗻 𝗿𝗼𝗹𝗲 𝘀𝗵𝗶𝗳𝘁𝘀 𝘁𝗼 𝗶𝗻𝘁𝗲𝗻𝘁, 𝗮𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲, 𝗮𝗻𝗱 𝗾𝘂𝗮𝗹𝗶𝘁𝘆 𝗰𝗮𝗹𝗶𝗯𝗿𝗮𝘁𝗶𝗼𝗻. Code production gets commoditized. Defining what good looks like, and building the evaluation loop an agent can self-correct against, becomes the differentiating skill.

𝟰. 𝗧𝗵𝗲 𝗲𝘃𝗶𝗱𝗲𝗻𝗰𝗲 𝗵𝗮𝘀 𝗮 𝗰𝗹𝗶𝗳𝗳. On EvoClaw, agents score above 80% on isolated tasks and at most 38% on continuous evolution across commit histories. The paper names the causes: context drift, error propagation, no model of technical debt, and tests that pass while semantics break.

𝟱. 𝗧𝗵𝗲 𝗴𝗮𝗽 𝗶𝘀 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴, 𝗻𝗼𝘁 𝗽𝗵𝘆𝘀𝗶𝗰𝘀. The author attributes the drop to context management, memory architecture, and verification. Those are buildable.

That last point is the whole reason CDE exists. Information architecture keeps context coherent across sessions. Quality gates catch error propagation before it compounds. Verification is the harness, not an afterthought. The model is the reasoning engine. We build the system around it.

Link in the comments.

09/02/2026

𝗔𝗜 𝗱𝗶𝗱𝗻’𝘁 𝗺𝗮𝗸𝗲 𝘁𝗵𝗲 𝘀𝗼𝗳𝘁𝘄𝗮𝗿𝗲 𝗹𝗶𝗳𝗲𝗰𝘆𝗰𝗹𝗲 𝗳𝗮𝘀𝘁𝗲𝗿. 𝗜𝘁 𝗺𝗼𝘃𝗲𝗱 𝘁𝗵𝗲 𝗯𝗼𝘁𝘁𝗹𝗲𝗻𝗲𝗰𝗸.

Addy Osmani just published his takeaways from Google’s new whitepaper on the AI-era software lifecycle. Five points worth your attention:

𝟭. 𝗔𝗻 𝗮𝗴𝗲𝗻𝘁 𝗶𝘀 𝗮 𝗺𝗼𝗱𝗲𝗹 𝗽𝗹𝘂𝘀 𝗮 𝗵𝗮𝗿𝗻𝗲𝘀𝘀. Roughly 10% model, 90% harness. Teams have jumped from outside the top 30 to the top 5 on Terminal Bench by changing only the harness. Most agent failures are configuration failures. Debug the harness first.

𝟮. 𝗖𝗼𝗻𝘁𝗲𝘅𝘁 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 𝗱𝗲𝗰𝗶𝗱𝗲𝘀 𝘆𝗼𝘂𝗿 𝗯𝗶𝗹𝗹. Static context is reliable and paid for on every call. Dynamic context loads on demand. Where you draw that line is an architectural decision. Version it like code.

𝟯. 𝗩𝗲𝗿𝗶𝗳𝗶𝗰𝗮𝘁𝗶𝗼𝗻 𝘀𝗲𝗽𝗮𝗿𝗮𝘁𝗲𝘀 𝘃𝗶𝗯𝗲 𝗰𝗼𝗱𝗶𝗻𝗴 𝗳𝗿𝗼𝗺 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴. Tests for the deterministic parts. Evals for the rest, on both output and trajectory. Set the bar at the eval, not the demo.

𝟰. 𝗧𝗵𝗲 𝗹𝗶𝗳𝗲𝗰𝘆𝗰𝗹𝗲 𝗰𝗼𝗺𝗽𝗿𝗲𝘀𝘀𝗲𝘀 𝘂𝗻𝗲𝘃𝗲𝗻𝗹𝘆. Implementation drops from weeks to hours. Requirements, architecture, and verification stay slow because they’re judgment work. Spec quality is the new constraint.

𝟱. 𝗩𝗶𝗯𝗲 𝗰𝗼𝗱𝗶𝗻𝗴 𝗶𝘀 𝗰𝗵𝗲𝗮𝗽 𝘁𝗼 𝘀𝘁𝗮𝗿𝘁 𝗮𝗻𝗱 𝗲𝘅𝗽𝗲𝗻𝘀𝗶𝘃𝗲 𝘁𝗼 𝗿𝘂𝗻. Token burn, maintenance tax, security cleanup. Past the crossover it can cost 3x to 10x more per feature. Context and model routing are financial levers, not just technical ones.

The line we keep coming back to: AI amplifies whatever engineering culture it lands in. Specification, verification, and the systems that hold them together. That’s where the work is now.

Full piece: https://oreillyradar.substack.com/p/the-new-software-lifecycle

By Addy Osmani

06/18/2026

Thought provoking piece 👇

06/16/2026

One of the questions you will often hear in the years to come, answered in a clear, concrete, and hype-free way.

Enjoy!

A field guide for engineers: the loop, the autonomy spectrum, the harness, and demo-to-production reality, with working code.

06/01/2026

Agentic Software Engineering — Pulse #2. Three things that landed this month:

• You can now steer a coding agent's behavior from inside the model — drift is a measurable signal you can correct on the fly.
• More compute isn't better. Accuracy peaks at a middle amount of spend, then flattens — and the agents can't predict their own usage.
• "An AI reviewed the code" often means no human actually did. Most agent-written changes get merged with little real review.

Read the full pulse 👇

https://open.substack.com/pub/aienhancedengineer/p/agentic-software-engineering-field

05/25/2026

05/12/2026

At AIEE, keeping up with 𝗮𝗴𝗲𝗻𝘁𝗶𝗰 𝘀𝗼𝗳𝘁𝘄𝗮𝗿𝗲 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 is part of the job.

Written notes and mental models were not enough. The field moves faster than any manual tracking system can handle.

So we built a 𝗽𝗶𝗽𝗲𝗹𝗶𝗻𝗲.

A verifiable ingestion and processing system with a single design goal: 𝗲𝗱𝘂𝗰𝗮𝘁𝗲 𝘂𝘀 𝗮𝘀 𝗳𝗮𝘀𝘁 𝗮𝗻𝗱 𝗽𝗿𝗲𝗰𝗶𝘀𝗲𝗹𝘆 𝗮𝘀 𝗽𝗼𝘀𝘀𝗶𝗯𝗹𝗲.

It tracks papers, anchors concepts to primary sources, ranks findings by a composite signal (key insights, cross-field reach, empirical claims, recency decay), and produces a structured digest the team reads before shipping anything.

We're calling the output a 𝗙𝗶𝗲𝗹𝗱 𝗣𝘂𝗹𝘀𝗲. And we're sharing it openly.

𝗣𝘂𝗹𝘀𝗲 #𝟭 is live. It covers the state of the field, the concepts crystallizing, the patterns stabilizing, and the top 7 papers worth your attention right now.

The headline finding: 𝗮𝗴𝗲𝗻𝘁 𝗯𝗲𝗵𝗮𝘃𝗶𝗼𝗿 𝗶𝘀 𝗮 𝘁𝗮𝗿𝗴𝗲𝘁 𝗼𝗳 𝗱𝗲𝘀𝗶𝗴𝗻, not just an emergent property of the model. Five converging papers from Jan-Apr 2026 make this case independently, and it changes how you should think about 𝗵𝗮𝗿𝗻𝗲𝘀𝘀 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴.

If you're building with or around coding agents, this digest was made for you.



https://open.substack.com/pub/aienhancedengineer/p/agentic-swe-field-pulse-1

2026-05-10 · State of the field, the concepts crystallizing, the patterns stabilizing, and this pulse's top 7 papers.

05/04/2026

“Which MCPs and tools do you use in your Claude Code setup?”

We have been asked this question repeatedly these last few months.

This article is the full answer, link in the comments! 👇🏽