Coding Agents and the Future of Software Engineering
Simon Willison emphasized that experienced software engineers remain essential for directing coding agents, leveraging domain knowledge and long-term context that agents lack.
Someone has to tell the coding agents what to do – someone with strong domain knowledge combined with deep understanding of how software gets built
Experienced software engineers are ideally suited to that role
— Simon Willison (@simonw) January 20, 2026
I'll believe it when I see it – personally the more time I spend with coding agents the less I'm scared for my own career, getting the best results from these things takes an enormous amount of experience and skill
— Simon Willison (@simonw) January 20, 2026
He praised Addy Osmani’s essay on Jevons Paradox, arguing AI tools lower code production costs, spurring more ambitious software rather than fewer developers.
This is really good, and worth reading in full – it's the best articulation I've seen yet of the idea that driving down costs involved in producing code will increase demand for software and let us take on much more ambitious projects https://t.co/wnQqKUELnm
— Simon Willison (@simonw) January 20, 2026
Surprising number of people saying they think this was written by an LLM – it appears that it's getting hard for people to differentiate between LLM generated text and good, well structured and well argued prose!
— Simon Willison (@simonw) January 20, 2026
Willison advocated letting agents like Claude Code experiment in safe environments, such as fixing Ansible issues, over relying solely on StackOverflow.
Did you try running Codex or Claude Code in an environment with Ansible available and having them hack away at the playbook until they figured out how to get it to work?
— Simon Willison (@simonw) January 20, 2026
Omar Sar highlighted Claude Code’s foundational capabilities for impressive workflows without over-optimization.
You don't need a crazy setup in Claude Code.
With just the foundational blocks, you can already do impressive things with Claude Code.
Stop overoptimizing around every tiny little feature/setting.
Aim instead for compounding workflows that fit your needs.
Good overview! pic.twitter.com/157LFH2jZz
— elvis (@omarsar0) January 20, 2026
Levelsio demonstrated Claude Code automating tasks while he flew a simulator on autopilot.
✈️ I finally figured out how the Airbus A320's autopilot works so I can set a whole flight plan and it just flies itself
So now I can fly my plane on autopilot while my Claude Code codes on autopilot too 😂
Just need Ralph Wiggum to completely remove myself here pic.twitter.com/2NFoL3S7DW
— @levelsio (@levelsio) January 20, 2026
AI-Generated Media and Creative Tools
Cocktailpeanut explored AI’s potential to create extreme influencers as “necromancers” unbound by real-world limits, evolving social media dynamics.
Social media's race to the bottom had a bottom because someone had to actually DO it or BE it. AI removes that constraint.
Yesterday's influencers are warriors (willing to do anything) or wizards (naturally gifted).
Tomorrow's influencer's are necromancers, spawning infinite…
— cocktail peanut (@cocktailpeanut) January 20, 2026
He showcased “Evolution Diffusion,” generating a human pianist from a dog video using LTX-2 and WanGP, including a whimsical “Bearman” transformation.
The Bearman https://t.co/cORmEZTlv7 pic.twitter.com/9Ec2xEOih3
— cocktail peanut (@cocktailpeanut) January 20, 2026
AIWarper highlighted Nano Banana PRO generating trendy outfits from a profile picture via Gemini 3.0 Flash prompts.
I didn't describe these outfits. Gemini 3.0 Flash did.
I simply dropped my profile picture in and asked for 4 ultra trendy hipster inspired outfit descriptions.
Nano Banana PRO did the rest.
Fit #2? Banger pic.twitter.com/1FeQXEvsx8
— A.I.Warper (@AIWarper) January 20, 2026
He demonstrated one-button AI filmmaking and fine-tuning LoRAs on Klein 9B outperforming base mega models for styles.
Just press a button
It’s that simple!
— A.I.Warper (@AIWarper) January 20, 2026
Another prime example of fine-tuning versus just using a mega model like Nano Banana PRO.
Both bottom images are from Klein 9B with a fine-tuned LoRA.
Bottom left uses the distilled checkpoint (no CFG, 4 steps).
Bottom right uses the base checkpoint (CFG 4, 20 steps). pic.twitter.com/74Fh53eL87
— A.I.Warper (@AIWarper) January 20, 2026
Additionally, he noted ElevenLabs Scribe v2 for voice changer and dubbing improvements.
Someone was recently asking me about voice 2 voice conversion improvements.
This one is worth a shot.
Also might be useful for LTX 2 dialogue improvements or if you want a consistent voice https://t.co/ISrqHorRcP
— A.I.Warper (@AIWarper) January 20, 2026
AI Agents Convergence and Open-Source Projects
Emad Mostaque observed AI agents converging due to near-zero costs for logic and data movement.
All AI agents are basically converging to the same thing with the cost of logic and moving bits intelligently moving to zero https://t.co/yoRLUM7EyV
— Emad (@EMostaque) January 20, 2026
He congratulated Eigent AI on 10k GitHub stars, part of CamelAI’s ecosystem including OWL, advocating focused OSS development.
Congrats!
Out of curiosity why didn’t you just rename OWL to Eigent and refactor code?
— Emad (@EMostaque) January 20, 2026
Cocktailpeanut critiqued Overworld’s local-first 60fps world model release.
bro forgot he's building a local-first world model https://t.co/9uD5gBdESP pic.twitter.com/jJdxh9Ukzt
— cocktail peanut (@cocktailpeanut) January 20, 2026
LoRA Training and Fine-Tuning Workflows
Ilker from FAL shared a detailed guide on training LoRAs for specific tasks like spritesheets, explaining why they’re useful despite advanced base models (e.g., cost/speed gains of 4-5x), data needs (e.g., 100 images for simple cases), paired images for editing (start/end suffixes), captioning, steps/learning rate balance, and inference.
As I promised yesterday, I'll briefly explain LoRA training and share a workflow I made so you can do it quickly.
First, let me answer a very common question:
'Why train LoRAs when we have such advanced models?'
Even though we have incredibly advanced models now (like NBP),… pic.twitter.com/AQ5MQ0zjxs
— ilker (@ailker) January 20, 2026
Enterprise AI Best Practices and Long-Task Agents
Ethan Mollick highlighted organizations adopting prior AI practices (frontier access, training, champions, prompts) but needing reinvention for agentic skills enabling long tasks.
Organizations have started catching up to the prior best practices for employee AI use (access to frontier models, training, champion programs, prompt libraries)…
…except almost all of this now has to be re-invented for a world of agents driven by skills that do long tasks.
— Ethan Mollick (@emollick) January 20, 2026
Model Limitations in Practical Use
Mollick criticized Gemini 3’s chatbot for inconsistent file delivery and code execution, making it less useful for real tasks compared to ChatGPT/Claude.
Not to repeat this, but the fact the Gemini chatbot can't seem to deliver files (or even consistently run code) is a huge gap compared to ChatGPT or Claude.
It makes a very smart model (Gemini 3) much less useful, especially for people trying to get AI to do real tasks and work. pic.twitter.com/FVbrAs40ks
— Ethan Mollick (@emollick) January 20, 2026
AI Creativity in Constrained Game Design
Mollick showcased Claude Opus 4.5 creating a coherent, fun story-driven game under absurd constraints (slider: Maximum Potato to Formalware; dial: Monet to Drive-Thru), with hand-drawn art, questioning human superiority.
Opus 4.5: "you need to build a game that is coherent, fun and story driven. There are precisely two controls. One is slider, in which one side is labelled Maximum Potato and the other is labelled Formalware, there are four positions for the slider. The other is a dial that goes… pic.twitter.com/ZrRv5sntKL
— Ethan Mollick (@emollick) January 20, 2026
Audio-Conditioned Video Generation
Justine Moore (a16z) demoed LTXStudio’s new feature for lip-synced videos from uploaded/generated audio (via ElevenLabs), predicting impact on character voices/AI influencers; tested with pet animations.
We've got a new audio-conditioned video model 👀@LTXStudio now lets you upload or generate audio, and then creates a lip-synced video.
This is going to be HUGE for consistent character voices + AI influencers.
I tested it w/ an animation of my pets. More on how to use it 👇 pic.twitter.com/UDKqpE9Cin
— Justine Moore (@venturetwins) January 20, 2026
AI Agents as Personal Gatekeepers
Moore proposed AI “bouncer” agents for email inboxes, requiring sender agents to justify entry with subscription proof, extending to X DMs for spam filtering.
My email inbox is so full of spam and newsletters I didn’t sign up for.
I want AI agent to be the “bouncer” and make the sender’s agent make the case for why the email should be let in.
Ideally with proof that I actually subscribed!
— Justine Moore (@venturetwins) January 20, 2026
Vibe Coding and Mobile AI Coding Tools
Riley Brown promoted vibecodeapp for easy Claude Code/Codex/Gemini CLI use on phones, with one-click exports; contrasted with Cursor’s browser integration, already live.
By the way… the easiest way to use Claude Code, Codex and Gemini CLI on your phone is on Sandbox dot dev.
Almost forgot we released this feature as a mini side experiment when testing how to run these models in a sandbox.
This product is still availible to all @vibecodeapp… pic.twitter.com/0ysGQMGSuA
— Riley Brown (@rileybrown) January 20, 2026
Great idea, all of it is live on @vibecodeapp. https://t.co/Q24JCCA0xX pic.twitter.com/B9JxyFYU3w
— Riley Brown (@rileybrown) January 20, 2026
Positioning as the “AI Expert”
Machina (EXM7777) urged becoming “the AI guy” in networks, exploiting perception gaps where clients assume broad AI skills for problem-solving, regardless of depth.
if you haven't developed a specific high-value skill recently… just become the AI person in your circle…
seriously, the market doesn't care about your technical depth right now, they care about solving problems
and AI is so obscure to most people that if you position…
— Machina (@EXM7777) January 20, 2026
AI Development Urgency and Economic Pressures
Machina warned of an AI “race against time” as ecosystems near fuel exhaustion, predicting price hikes/quality drops; build now while access is cheap.
you're in an ai race whether you realize it or not…
not against other people, against time
this whole ecosystem is running out of fuel and when reality hits, subscription prices will explode or models quality will go to shit
you have access to the most powerful tool in human…
— Machina (@EXM7777) January 20, 2026
Marketing Context Profiles with AI
Machina detailed using AI for marketing by compiling competitor profiles (copy, funnels, ICP) into JSON for LLM campaigns, avoiding generic output.
here's how to use AI for marketing:
build marketing context profiles
>find the best offers in your niche
>study their copywriting, voice, funnels
>analyze what specific mechanism they use, their language, the ICP
>compile everything into detailed JSON profilesthen feed these…
— Machina (@EXM7777) January 20, 2026
OpenAI Codex Adoption Surge
OpenRouterAI charted rapid growth in OpenAI Codex usage, topping model rankings.
OpenAI Codex growth 👀
Top 5 models below pic.twitter.com/YL5fhu0rQs
— OpenRouter (@OpenRouterAI) January 20, 2026
LLM JSON Response Healing
OpenRouterAI updated Response Healing with smarter error guidance for common failures (arrays/strings), plus docs.
NEW: Smarter error messages for JSON healing ❤️🩹
When models return arrays (68% of failures) or strings instead of objects, you'll now get guidance:
– Arrays: Use json_schema {"type": "array"} or wrap in an object
– Strings: Explicitly request JSON in your prompt https://t.co/tk8wClIvne— OpenRouter (@OpenRouterAI) January 20, 2026