Observation from the AI coding trenches.
There are two emerging trends (or observations) that I’m seeing constantly across a number of startups and a growing number of enterprise projects. In fact, they are so repetitive that I’d categorize them as stable patterns at this point.
1. Human Code Review Going Away
Throughout the 2024–2025 period I was a staunch proponent of reviewing just about every line of code that AI generated. Back in late 2023 and early 2024 that was an absolute necessity as the quality of generated code was simply lacking.
In 2025 I started noticing, at least on a personal level, that manual code reviews were becoming more of a formality. You are still doing them mostly because you’ve been doing it your entire professional life.
In the last six months my opinion has been shifting rapidly. Today, AI agents can do much better code review in just about every use case: security and compliance review, performance, reusability and componentization, multithreading, distributed programming, SQL/NoSQL backends, front-end code, and even hardware-centric device code - you name it. More importantly, agents can do it more reliably and much faster.
Have I abandoned manual review entirely? No. Skills, sub-agents, and hook setups are not trivial, and the cost can still be prohibitive. But the writing is technically on the wall, and 2026 is likely the year when widespread manual code reviews for AI-generated code start diminishing rapidly.
2. AI Coding on Greenfield vs. Brownfield Projects
Yet another stable pattern emerging is just how different the agentic coding adoption rate is between new projects and existing ones.
While new projects have a relatively easy adoption curve - with new org-wide context-management tools coming online almost weekly - introducing spec-driven agentic coding into existing teams and projects is an absolute pain.
Personally, I was not expecting this at all. Yes, the differences are obvious, and existing projects clearly require more upfront investment and refactoring. But the reality is much harsher: it is extremely hard to introduce spec-driven agentic coding into an in-flight enterprise project. Frankly, it is nearly impossible to do so without major disruption. It wrecks existing team structures, CI/CD pipelines, timelines, project management processes, and relationships with business stakeholders.
It’s an uncomfortable truth about just how disruptive this AI transformation is for the software development industry at large.
Almost everything has been said about Anthropic vs. US government.
- If this “supply chain risk” decision stands and is enforced, Anthropic is done.
- In that scenario, acquisition is the most likely outcome.
- It’s INCREDIBLY frustrating that one person’s political views and sense of moral entitlement can materially damage what is arguably the best AI company in the world right now.
- Where the f*ck is Anthropic’s board? Decisions of this magnitude are supposed to be made at the board level - not unilaterally by the CEO.
- The fact that OpenAI signed a deal just hours later (with common-sense restrictions that were also offered to Anthropic) tells you everything you need to know about Anthropic’s negotiating position and leverage.
- Broad political or policy disagreements belong at the ballot box, not embedded into adversarial contract negotiations with the US government.
- The whole situation is getting absurd on both sides - but at least the US government’s position is clear and reasonable: no private company gets a kill switch over US military operations.
The AI Intelligence Approach Took a Hard Turn.
Just 18–24 months ago, most of what we called AI intelligence - reasoning, cognition, problem-solving - was almost entirely attributed to LLMs. There was a broad expectation that the next SOTA models (e.g., GPT-5) would deliver a step-function improvement in reasoning itself.
Fast forward to today, and the center of gravity has shifted decisively toward agents.
Modern agents (Claude, OpenClaw, etc.) learned how to recursively decompose large, ambiguous problems into smaller units that current-generation LLMs can already handle well. As a result, expectations from LLMs have narrowed to two core capabilities: (a) reliably following instructions and (b) reasoning effectively over bounded, small-scale tasks.
Agents, meanwhile, are evolving at a different pace. Tooling (MCP), skills, sub-agents and A2A coordination, memory, and execution frameworks are expanding almost weekly. In just the last 30 days, we’ve seen the launch of a global, planet-scale, persistent, self-regulating memory plane for AI agents (Moltbook). OpenClaw’s growth has been equally striking.
Agent intelligence now clearly exceeds what any single LLM can deliver - and, more importantly, these capabilities compound in a strongly non-linear way.