Predictions

After listening to the excellent LLM predictions for 2026 podcast episode from Simon Willison and Oxide, I decided to try a few of my own. If nothing else, looking back will provide an amusing respite from serving our robotic overlords next year.

1. Coding interviews will change. Coding interviews will stay the same.

   Starting off nice and noncommittal on a topic that I haven’t seen much discussion on yet. It’s clear that the role of engineers is changing significantly and so the way we measure people in an interview setting will need to too.

   As with the slow move away from people trying to emulate 2000s Google Whiteboard Interviews though, most people will just keep testing manual coding as an effective proxy in the near future.

2. No new (human readable) languages

   One of the main drivers of people creating new programming languages has always been to make ones that are easier for humans to write, read and understand. With this gone, fewer new languages are created and there’s little uptake for those that are. Whether agent-specific languages emerge is something I’m less sure on.

3. “Modes” (Plan etc.) disappear

   This falls into a general bucket of techniques (MCP servers, Ralph Wiggum loops, masses of markdown files) that I think will seem quaint in hindsight. The harnesses will quickly improve and be able to respond more appropriately and interactively based on input (does this seem like a question?) and confidence level in jumping straight into a solution.

4. Agent capabilities commoditise, big players compete on scaffolding and UX

   We’re already seeing this to some extent, but I think the leading models will become good enough as to be indistinguishable from one another, forcing the major players to find other axes to compete on such as multi-agent orchestration. Simultaneously, open-source models and harnesses will become Good Enough for many use cases.

5. Context window size is no longer a blocker, attention still an issue at high context volume

   Even when you can fit an entire codebase into the model’s context, problems will remain with not reusing existing methods and patterns in some cases.

6. Mid tier of open source fades

   Large projects remain, niche / personal projects remain, but a combination of ease of vibe-coding your own solutions (demand side) and reduced need for community contributions (supply side) means we see fewer moderately successful OSS projects.

7. Baseline of design rises, good human-led design is still at a premium

   Similar to what we saw around the initial Bootstrap era, agentic tools are letting non-designers get much better baseline results. I think the Indigo Problem will improve, but premium products will still lean heavily on human design.

Discuss

Let me know what you do / don’t agree with via the Hacker News post!