AI coding roundup
The signals from the noise
19th Feb 2026
First, a PSA
- A lot of online stuff is like customising excel (btw it’s an esport)
- The purpose of these tools is that anyone can code now
- Sure, you can run 12 excels at the same time overnight with custom macros
- But the point isn’t to get better at excel, it’s to use it to solve important business problems
- e.g., if you ship a key project that creates shareholder value early then it likely won’t matter how (whether you filled out your Jira tickets, have high test coverage, had 90 agents in parallel, etc.)
- The default settings and best model (currently probably opus 4.6) is just fine (i.e., most of the way there)
Stuff
- New best models
- $$$: Opus 4.6
- $$: Sonnet 4.6 or Codex 5.3
- Auto AI code review (very useful)
- e.g., GitHub Copilot with
.github/workflows/copilot-instructions.md - AIs are pumping out lots of code so need to sample it enough
- Give evidence that this PR solves the problem e.g., screenshot from a staging test
- This reduces the burden on human reviewers
- Also ensure your PRs are right and consider making smaller
- e.g., GitHub Copilot with
Tips
- Become good at code review (similar) - develop verification of whether the code solves the problem
- What is taste in software - flexibility, design (i’m exploring this more)
- Multi-task with background agents - worth experiment when comfortable
- Use plan mode - can get better results for complex stuff
- Ideas on how to use LLMs (similar) - lots for different teams and roles e.g., k8s debug with screenshots, explain a codebase, cross-language translation, refactoring, make tests, improve existing code, debugging (paste in traceback)
- Make skills after LLM/you done it first
- Performance can degrade when context window >50% full
- Easily add parallelisation
- Security tips - careful with browser use, using stuff from internet, what commands you approve
- Try and have loops e.g., with tests
- Ensure every repo has a
AGENTS.md, that update regularly - Try it as a research assistant to “deep research” papers for ideas to implement
- Take the time to learn and try out stuff
- Chat with colleagues about how they use it
Opinions
- ICs now agent managers (similar) - popular i.e., architects now
- As implementation cheaper, do tasks currently unable to - nice - what never had time to do?
- Advice for junior and senior developers (similar) - learn, adapt, and focus on what agents can’t do (understanding, creativity, design, trade offs, business context, domain knowledge, communication (both with agents and humans), etc.)
- What to do while for agents - I often think about the problem/solution, what it’s doing, or multitask other stuff e.g., read, email, slack, sometimes fire off another agent (plan to experiment more with this)
- Downsides of vibe coding - I’m often more jumbled / not in flow
- How measure productivity - many metrics miss the point i.e., solve important business problems
- Mid 2025 study found devs less productive with AI, recent anthropic study found 25% (not 10x) but learnt less - I guess might be different now tools are better
- What speed ups are realistic - i’m interested in results for real world tasks, who knows, i guess we’ll see in speed of real things being shipped, GDP, etc.
- Careful with cognitive debt (similar) - really try and understand stuff
- Context switching is tiring - yep
- Avoid skill atrophy - I use the LLMs as a 1-to-1 tutor for me (very valuable!)
- Me
- I find using agents fun. I don’t really miss coding by hand
- Though it’s challenging keeping up, context switching, and adapting to big changes / uncertainties
- I liked this tweet