# /NEW 2026 Highlights
## Highlights and Takeaways from /NEW 2026
### Day 1 Highlights
**AI is like chocolate — but won't make everything taste better (Liz Fong-Jones)**: AI doesn't automatically make things faster; it enables parallelism. Below-average practitioners might see their floor raised; above-average ones may find it drags quality down. The principal engineer who publicly told Liz she was "shipping slop" is the point — it took a human to catch what the AI code review agent missed.
**80% of delivery work is the same everywhere (Simone Bennett)**: Her 80/15/5 rule — 80% of delivery work is identical across projects, 15% looks bespoke but is just unconfigured, 5% is genuinely custom. Most teams have inverted that in their heads, which is why senior engineers are perpetual bottlenecks. "_If you can't be replaced, you can't be promoted._" The fix is unsexy: write it down, harvest the pattern, make it repeatable before the agents arrive.
**AI will isolate the water plant (Elena Scifleet)**: In her nuclear simulation, no AI model ever chose to de-escalate; all showed a consistent preference for **escalation**. In cybersecurity incident response, the same bias means an AI would correctly isolate a compromised water utility system, potentially cutting off public health infrastructure across an entire state. The contextual, ethical, and regulatory calculation that prevents that call is irreducibly human.
**Two words can corrupt an entire investigation (Dan Clements)**: One early agent in a pipeline saying "possible data breach" can anchor every subsequent agent to that framing. An investigation that should have resolved as a minor anomaly can inflate to 7–8 million tokens of compounding certainty about a breach that never happened. He called this cascading catastrophe — a failure mode specific to multi-agent chains that single-agent thinking doesn't anticipate.
**The data scientists refused to build it (Panel — Catherine Squire)**: At Culture Amp in 2022, the engineers with the deepest knowledge of the technology — who understood hallucinations, bias, and real model limitations — were the ones who considered an AI summarisation feature unethical. The resistance came from the people you'd expect to be most enthusiastic. The resolution required radical transparency and treating them as experts, not overriding them.
**The AI vampire accelerates whatever trajectory you're on (Navin Keswani)**: AI tools are engineered for maximum engagement and session length. They produce dopamine. If you have a propensity for burnout, AI takes you there faster. The fix is to design your own system rather than work in default mode; state your current state and goal before starting, pre-commit to a time limit while your judgment is still fresh, and close each session with a debrief. "_The more tired you are, the worse your judgment about whether you're too tired to think effectively._"
[D1 notes →](/posts/2026-05-27-newcastle-slashnew-2026-day-1/)
---
### Day 2 Highlights
**The 84% nobody talks about (Jovana Dunisijevic)**: Writing code is only 16% of the software development lifecycle. AI tooling is almost entirely focused on that 16%. The real leverage — requirements clarity, architecture decisions, governance, alignment — is the other 84%, and barely anyone is asking AI to touch it.
**The 500-day problem (Nick Williams)**: The LinkedIn MVP-in-a-day stories are often real. What's missing is the day-500 follow-up. Organisations with genuine long-term AI success had proper delivery infrastructure before they started. AI amplifies whatever foundation already exists, good or bad.
**You own the code (Michelle Sandford)**: When a coding agent makes a change, your name is on the commit. AI changes how code gets written, not who is accountable for it. Approving a pull request you don't understand used to be a constraint; now it's a choice.
**Intelligence vs. wisdom (Jovana Dunisijevic again)**: Intelligence identifies patterns; wisdom chooses direction. As AI intelligence becomes more abundant, wisdom becomes scarcer and more valuable. Organisations are investing almost entirely in the former.
**The talent pipeline is the oversight problem (Charlotte Fleming)**: Cutting junior hiring looks clean on a spreadsheet. The cost shows up 3–5 years later as a senior skills shortage. The deeper issue: human oversight of AI only works if the humans have enough experience to actually oversee, which takes years to develop and can't be assumed to be there.
[Day 2 notes →](/posts/2026-05-28-newcastle-slashnew-2026-day-2/)
---
### Takeaways
**The infrastructure argument won the conference.**
Multiple speakers converged on the same point from different angles. Simone Bennett argued you must codify delivery patterns before agents arrive. Nick Williams demonstrated experimentally that soft guardrails produce high variance; hard platforms produce consistency and lower cost. Charlotte Fleming's DORA data showed that AI adoption without mature delivery practices actually reduces throughput. Liz Fong-Jones noted that the teams genuinely succeeding at day 500 were the ones that already had standards and guardrails in place. Nobody coordinated it. The conclusion was the same: AI amplifies whatever substrate it runs on, good or broken.
**The 84% problem kept resurfacing.**
Jovana Dunisijevic's framing that code generation is 16% of engineering work was the sharpest version of a claim that ran through half the talks. Charlotte Fleming mapped the same gap empirically: what AI does well and what developers actually want help with barely overlap. Dan Clements built his multi-agent SecOps architecture precisely to reach the investigation work that never surfaces in an alert queue. The conference largely agreed that the current focus on coding assistants is the least interesting part of what AI can do.
**Accountability without accountability structures is meaningless.**
Elena Scifleet, Dan Clements, Michelle Sandford, and Charlotte Fleming all addressed this from different directions. Michelle's formulation was the most direct: your name is on the commit, not the agent's. Dan's was the most operational: agents at NIB cannot perform destructive actions, full stop; chain-of-thought logging is mandatory; agents operate under distinct identities. Elena flagged the governance gap nobody is closing: when an AI-driven IR response gets scrutinised by regulators or in court, current frameworks cannot answer who is responsible. The speed at which AI acts makes the radius of a wrong decision larger, not smaller. More automation requires more governance, not less.
**The talent pipeline is a slow-moving crisis that is already in motion.**
Charlotte put numbers on a concern several speakers raised. 73% of organisations have reduced junior hiring over two years. Senior capability takes five to seven years to develop from a junior starting point. The shortage will appear as senior wage inflation two to three years after the hiring freeze, as it has after every previous contraction. The same people needed to provide AI oversight are the ones organisations keep treating as the first line item to cut.
**AI accelerates whatever trajectory you are already on.**
This was stated explicitly by Navin Keswani on burnout, by Charlotte Fleming on delivery infrastructure, by Jack Skinner through the runway analogy, and by Liz Fong-Jones through the mole metaphor. The pattern is consistent: the tool amplifies. If the underlying system is well-designed, AI compounds the value. If it is broken, misaligned, or dependent on institutional knowledge that was never written down, AI compounds those problems faster and less visibly than before. The corollary is that improving the underlying system is not optional preparatory work. It is the work.
**"Humans in the loop" is the wrong frame, but the right instinct.**
Jovana's opening complaint about the phrase became the conference's running thread. Several speakers made essentially the same point: the question is not whether humans should be in the loop but which loop, at what stage, with what authority, and with enough knowledge to actually exercise judgment. Michelle described four distinct governance positions with different risk profiles. Dan drew a hard line around destructive actions. Elena showed that decisions involving public health, legal obligation, or crisis communication are not technical questions and cannot be automated. The conference consensus was not "humans first" as a slogan. It was a more specific claim: human judgment is irreplaceable at the decision gates that carry consequence, and those gates must be designed, not assumed.
## Gallery:
