AI Engineer World's Fair 2026 Roundup

RIP IDEs and Code Reviews?

Back in November at AIE Summit NYC, the marching orders were clear: terminal computer use and coding agents should be at the core of not only how you build products but how the products themselves should work. Unrestrained agentic behavior within sandboxes, going crazy with bash and writing python. Shortly thereafter, Claude Cowork dropped which validated this approach to product and knowledge work.

For the world's fair, I was curious what the big takeaway would be. Early hints on arrival suggested "loops": instead of babysitting agents interactively we orchestrate them in such a way that they just plug away and derive their own prompts from the environment, perhaps ticket queues, slack, production telemetry, etc. This was prominent, particularly the idea of a "software factory" which is building these cathedrals of higher order loops, orchestration, and verification so human in the loop is very high level.

However, there were more specific notions that struck me the most.

Lopopolo vs Zechner, the Z/L Continuum: Should developers be READING code any more

I thought we were still debating the idea of if developers need to use IDEs or ever write code by hand, but no, the debate is now whether we should be reading it at all. Lopopolo vs Zechner refers to a conference where two credible individuals argued the opposite back to back and both received enthusiastic applause.

The consensus seems to be "yes, for now" if you care about drift in your architecture, repo, dependencies etc. In fact, fading understanding of the structure of your repositories and atrophy of detailed technical know-how ranked as the #1 consequence of agentic AI in the conference survey.

However, there is a smaller subset (I count myself in this group) that believe if you wire up enough adversarial review/acceptance criteria/etc. in your software factory, are careful in the planning and research phase, and throw xhigh frontier models at the problem… then you don't. It's not like humans don't make a mess of codebases anyway, so why not just fix-forward anything that pops up? Or just skim PRs after the fact when appropriate? It may depend on the consequences and audience of the software but alas, this is the trajectory.

Turn the lights off when you leave?

This is related to the last one, but there is debate as to if it is currently possible to build a robust enough set of loops to just take humans out of the loop for the "Lights-out Software Factory." Maybe not yet, but this is the conversation… and it does seem like a tangible possibility and not just AGI hype.

Net-net: In 1 year we will not be writing, reading, reviewing, or even approving code. Let that sink in.

Define a goal and let the agents figure it out—the end of profiling, tuning, and iteration

There's a few different elements to this which I will describe, but they all share a common idea: define a goal and let the agent hammer it into shape.

The coarsest version of this is /goal in claude code/codex. "Keep working until all the tests pass and the report loads in under 100ms". Or "keep improving performance, don't give up" then come back in the morning… the Ralph loop.

The more refined paradigm shift to internalize is Autoresearch. If you have code and/or a prompt, an eval that scores the outcome (e.g. perhaps its a performance or accuracy score), you can let the agents brute force options for you instead of ideating and iterating yourself. It becomes a search problem, and the candidates come from the models' own creativity and your cues.

Weco gave a great talk where they used this approach to win OpenAI Parameter Golf (tuning a small 15mb model) but it can be used for anything. Weco has released a harness and web GUI for this process that creates a tree of candidates, pruning lower scoring variants and expanding on successes to get better and better outcomes. There's a related approach called GEPA / optimize_anything which features breeding variants to optimize prompts or code.

Claude code dynamic workflows, RLMs (recursive language models), DSPy are all complementary techniques: give the model freedom to write its own code to reach a defined goal, then layer one of these optimization approaches like GEPA or AIDE to not simply accept the first attempt but rather breed or search for the best variant. These frameworks force a little sanity and structure at the seams to allow this process to work.

Time to elevate and apply these approaches to any chunk of work where there's a notion of accuracy or performance. It is a shift: you can half-ass the first version of a prompt or implementation or possibly not provide one at all.

GLM 5.2 is looking like a big deal

Heterogeneous model architectures are gaining steam. Whether it was the local crew thirsty to run GLM or folks crowding the Z.ai booth at the expo asking which inference provider gives the best single stream latency, this model was definitely the talk of the conference. One clever idea I really liked: passively log production trajectories and test out of band which parts of the workflow can be run on GLM without quality loss instead of evals directly.

Other

  • RLMs and recursive coding subagents: don't have to manage context rot if you let the model divide and conquer large tasks recursively and aggregate the results
  • Developers swear by switching from text to voice using wisprflow.ai or handy.computer. 3x the speed of WPM of typing, works anywhere, and understands filenames and technical terms.
  • Dedicated PR review tools like Greptile are still highly popular in addition to standard adversarial reviews (/review) within your dev flows
  • Waymo felt solved after 4 rides, Unitree staff were flanked by androids, yet still seems like physical AI is coming-soon story for the generalists?
  • Local: Gemma 3.5 4B gives you GPT-4o on your mobile phone which is crazy, but still seems like we are 1 generation away from Nirvana on local frontier power as DGX spark and Mac Pro both have shortcomings. You can have Nirvana today for $100k: it was cool to see a GB300 DGX Station in person.
  • "We have entered the world of good, fast, cheap - pick 3" - Thariq Shihipar, Anthropic
  • Conference survey: 17% have non-devs shipping full features to prod and 1/3 call software engineering a solved problem!
  • OpenAI 5.6 sol @ 750 t/sec on Cerebras - need to try this ASAP

Final Thoughts

I am starting to see my role at this point to be a high-level babysitter, goal-setter, and final judge. It can be case by case how much I can get away with this (the stakes, how much effort I've put into the harness and loops, how lazy I've been defining the goals and acceptance criteria, the availability of signal), but this is the direction of travel.