In 2026, I wanted to return from management in a corporate environment to hands-on work in a dynamic environment. Joke's on me! LLMs have become mature enough to take hands-on work right out of my hands. At least, some frontier models, for much of the hands-on tasks, and yet I"m busier than ever.
It had been unthinkable just months previous.
In early 2026, the ways LLMs did more good than harm in my daily work were:
- Comparative research on test tooling
- Rubber ducking on architectural decisions
- Rubber ducking on Quality Engineering transformation
- Research on CI/CD tooling and configuration
- Research on emerging test approaches
Every month since May 2026, that has changed.
Increasingly, I've been able to let coding agents:
- Draft test plans based on ticket intent and risks (not just on the implementation and documentation), then independently write & run throwaway automation for functional verification and exploratory testing, then translate the test results into succinct but actionable findings and an overview of mitigated risks, while handing over decisions, ambiguity, and non-automatable scenarios to a human (me)
- Perform lightning fast analysis of large quantities of logging data
- Drive iOS & Android emulators and manage test app versions
- Perform detailed comparisons of browser rendering against designs, with annotated side-by-side screenshots
- Prepare complex test conditions while I'm fetching a cup of tea
Since developers also use coding agents, the amount of delivered work increases dramatically. That can only go right if the development lifecycle has lots of sensible quality gates and automation, just like you'd expect in fast-moving organisations even before LLMs were a thing:
- Fast, meaningful, reliable tests at every stage of development
- Safe, controlled production releases (feature toggles, canary releases)
- Excellent product discovery with well-defined specifications
- Systems thinking-based incident management
- An appropriate monitoring and alerting configuration
Quality Engineering in the age of agentic engineering, where an organisation previously releasing 10-20 pull requests to production per day might now release 100+ per day, requires the same quality gates, just assisted by LLM tooling. So rather than an entirely new way of working, it's the same way of working with fewer human touchpoints.
That might mean:
- Allow code reviews, ad hoc test automation, exploratory testing, and fix and retest cycles to be handled independently by coding agents
- For low risk tickets, that might be exclusively done by bots
- For medium or higher risk tickets, the bots do the first pass, and hand over the remaining risks to human judgment and investigation
- A bot helps humans write clear and actionable specifications, with assumptions exposed and tested
- Monitoring anomalies are identified and investigated by bots and fixes provided immediately if approved
- Bots help generate, heal, as well as prune unit, integration, API, and UI-level test automation
Let's see all of this become outdated before the end of the year. 😅
But what about technical debt ballooning because of low quality generated code?
Currently, shockingly, the internal quality of code generated by a frontier model like Opus 5.5 (at least up to 4-5 November 2026, when many reports appeared about a sudden decrease in quality) has become comparable to that of medior and in some respects senior engineers.
However, the sheer scale of changes does mean internal quality deteriorates at a faster rate than before. The counterpoint is that technical debt can also be tackled at a faster rate, if any effort is put towards it.
But when delivering that much more work, how much of it is actually valuable?
That's probably the trickier part. It's easy for anyone, now truly anyone and not just Senior Management and the Technology department, to get carried away with what's possible. Someone needs to have a firm grip on which changes help the organisation and its customers and which don't. If you don't have that someone, either become them or grab some popcorn for the vibocalypse.