Coding Agents Are Breaking Containment: The 2026 AI Stack, the Coding Wars, and the Next Frontier

At 1:02, Swyx had just returned from AIE Europe, where the conference runs roughly once a quarter. He said the track list is a map of his priorities: OpenClaw is the story of the last four or five months, just below it are harness engineering and context engineering as related topics in agents and RAG, and then there is a long tail of evergreen work such as evals, observability, GPUs, and LM infrastructure. Multimodality and generative media are also on the agenda, but the first three are what people are most focused on.

At 2:06, Jacob brought up Harrison Chase's tweet that the infrastructure around AI finally feels stable. Swyx said it is expensive to claim this time is different, but when you are writing code it is fine to make the call, and being right on the thesis matters less than figuring out how to monetize it. He noted that packaging integrations for agents has landed on skills—a markdown file with scripts attached—and he does not see how it can get simpler. If the consensus stack is LLMs with tools in a loop, a file system for retrieval, and skills, then some stability makes sense, though real-time elements, subagents, and memory will keep adapting; he would not stake his reputation on the stack staying fixed.

At 4:12, Jacob contrasted application companies with infrastructure companies. Application companies have an end customer that is relatively sticky, while infrastructure companies sell to developers who are pickier and can churn completely to the hot new thing. Swyx agreed and framed the tradeoff as the vertical-versus-horizontal debate in AI startups: an application company can act as an outsourced AI team, translating model capabilities into customer outcomes, which is robust to whatever changes in the engineering layer.

By 5:44, Swyx argued that the useful horizontal companies are mostly reinventions of classic cloud for the AI era, with sandboxes as the clearest example. Sandboxes are another form of compute, and the workloads are enormous even if the category is less glamorous than it sounds.

At 6:49, the conversation turned to whether companies should train their own models. Swyx described an agent-lab playbook: start with state-of-the-art models, specialize on a domain, and once you have enough workload and high-quality user data, train your own models to save on cost and latency. He also acknowledged a marketing bonus, but said there is real value: Composer 2 and SWE 1.6 show up in the top five on a model switcher even when they are not subsidized. The pattern is a reversal of the bitter lesson—bootstrap on large general models, then distill to smaller models for high-volume, low-variance workloads.

At 9:26, Swyx brought up custom chips as a force in favor of open models. Cerebras and other non-Nvidia hardware are pushing inference into thousands of tokens per second instead of less than a hundred, which changes the quality-cost tradeoff. Every 10x in speed can unlock a different usage pattern, and the investment cycle is likely multi-year, so he cautioned against dismissing it.

At 11:32, Jacob asked what changes when agents become the buyers of infrastructure. Swyx rejected the idea that agents are the ultimate rational developers: they are easily prompt-injected and tend to compound existing winners. He cited Malte Ubl's stat that 60% of traffic to Vercel's app/admin configuration is bots, mostly coding agents using a CLI or MCP.

At 12:33, Swyx laid out the practical implication: if it does not exist as an API that agents can use, it does not exist. Products should also work on the CLI, not just the UI. His deeper point was that agent experience should really be developer experience—good docs, a consistent mostly stateless API, and discoverability or progressive disclosure. He was skeptical of AEO-style gaming of chatbots, though he expected short-term wins to compound.

At 14:35, the discussion turned to how a new challenger competes with an incumbent like Supabase. Swyx pointed to Resend, which started in 2023 and was recommended by Claude in about 70% of unprompted requests for an email provider in a recent survey. The shortlist in a model answer is likely to be three names, not twenty, so semantic association and combined-use articles matter. He also expects better memory and personalization to eventually outweigh frequency of mentions.

At 17:38, Swyx sized the coding market. Anthropic and OpenAI have made coding a P0, and Anthropic is at roughly $2.5 billion in ARR from Claude Code, though how ARR is recognized is debatable. OpenAI's public number is not known but is around $2 billion, and Cursor is rumored to be around $2 billion. Claude Code had just celebrated its one-year anniversary.

At 18:39, Swyx pushed back on the idea that founders should bet on the empty space outside coding because coding is already 50% of Claude use cases. He compared it to Apple versus Google: if a market went from 10% to 50% in a year, why assume it cannot keep going? The mean-reversion bet is painful if the momentum bet would have worked. He also described the current phase as capability exploration: teams are rewarded for spending more tokens rather than spending less, so the sloppy token-maxing leaderboards are directionally fine because the industry is still underusing AI.

At 20:15, Jacob brought up Ryan Lopopolo, who spends a billion tokens a day—roughly $10,000 a day at market API rates. Swyx said most of that may be slop, but the person living at the edge will discover new capabilities first because they are running experiments others are not. He also noted that Anthropic is the premium, restrictive player while Codex and Gemini are heavily subsidized, so a capabilities explorer can do a lot on a $200 plan.

At 21:48, Jacob asked what the end state is and whether it generalizes to other markets. Swyx's base case is status quo: two big players and a long tail serving use cases the giants do not. A major change would require a shift in economics, brand, or value proposition, and he had not seen one in the last six months. Microsoft/GitHub, Meta's MSL, and Chinese labs such as Z.ai with GLM are all trying, and he flagged finance and healthcare as the next verticals the model labs will pursue. He also noted that both big labs are testing TAM expansion—Claude Code for finance and Claude Cowork, plus OpenAI's super-app push—while Cursor and Cognition stay focused on coding, leaving space for application companies that do the last mile.

At 26:26, Swyx called the scale mindboggling: xAI was IPOing at $1.2 trillion, and the big AI companies are not slowing down the way normal startups do after reaching a ceiling. Jacob said the unusual dynamic is that late-stage outcome bands are now enormous.

At 27:28, Jacob argued that later-stage AI companies have outcome bands so wide that every three months can be an existential event in either direction. Anthropic is the positive case: for a while it was unclear whether it would have enough capital, then coding hit at the right time, it had the perfect model, and it executed brilliantly. OpenAI is also crushing it, so being number two in coding is a champagne problem.

At 28:31, Jacob noted that many people think Codex is as good as or better than Claude Code. The surprise is consumer AI: ChatGPT had a first-mover advantage, and even if Claude and Gemini are comparable, people stick with ChatGPT because it introduced them to the experience. Swyx said ChatGPT is not growing anymore, but he sees that as a category-wide plateau in consumer AI rather than a share loss—coding is parabolic, while consumer AI has not figured out how to bring on many more daily active users.

At 30:04, Swyx said it is still early: ChatGPT is three-plus years old, Claude Code only one. In a high-temperature phase, loyalty to first movers is lower than in more stable markets. Still, the Claude Code lock-in surprised Jacob, who had expected enterprise APIs to be less loyal than consumer products; the enterprise product turned out stickier. Swyx also noted that Codex is still catching up in his personal experience, and that the missing question is why coding is concentrated in two names when Gemini and xAI are trying.

At 32:08, Jacob asked what the next magical product experience might be. Swyx guessed something around consumer agents, computer use, and a hybrid of the two; OpenClaw is a vision of things to come, and although OpenAI has the association, it has no guaranteed win. His thesis is that 2025 was the year of coding agents and 2026 is when coding agents break containment to do everything else. The logic is a transitive property: software eats the world, coding agents eat software, therefore coding agents eat the world.

At 33:41, Jacob said the vibes narrative overstates organizational shifts at OpenAI, which still has great products, a great team, and great models. He argued the ecosystem is incentivized to want two or three more strong model companies, so he would be shocked if the next six to twelve months did not bring at least a partial reversion toward a more equal vibe. Swyx agreed but pointed out that different stages matter.

At 34:41, Swyx said the neolabs have not really broken through in the past year. The question is how to grow diversity and choice. Nvidia is incentivized to broaden the model-provider platform, but Swyx quipped that Nvidia tries harder to build NeoClouds than to build labs—CoreWeave is in a much happier place than a neolab built on top of it.

At 35:44, Jacob asked whether foundation models will eat startup categories. Swyx said not really for very small startups: the downside is minimal because the worst case is being acquired into one of the labs, and competent execution is a job interview. Mid-size startups are different. He pointed to consolidation in AI infrastructure—observability and tooling companies being absorbed into data platforms—as evidence that the domain-specific playbook is being worked out.

At 36:47, Swyx said traditional SaaS should be more worried. He is living through it at his own company, which pays about $200,000 a year for event and sponsor management SaaS; he could probably spend $2,000 and build a custom version. The hard part is not the build but bringing the rest of the team along. He is the most technical person on his team, and he cannot make that decision alone.

At 38:21, Swyx described the widening disparity in AI enthusiasm inside companies. One side says less AI-native colleagues are close-minded and should just hold their nose and start talking to agents; the other says the vibe coder shipped an 80% solution in a weekend and left employees to pick up the rest. That debate is happening in every company. Swyx is hesitating to pull the trigger on replacing the SaaS, but expects to within a year, not five.

At 39:23, Swyx said SaaS is definitely getting squeezed, which raises the question of an AI-native system of record beyond Postgres or MongoDB. Convex comes up a lot, but the Firebase of AI apps does not really exist yet. He wants a system that supports rapid iteration first and scales later, rather than starting with older database technology.

At 39:55, Swyx recalled a dinner with Anthropic CPO Mike Krieger where the table went around naming what worried them most. Swyx raised bio safety, and the table reacted as if he had said someone could build a virus that wipes out half of humanity; Mike, sitting on Anthropic's Mythos model, knew the answer was security. Swyx also argued that Anthropic's release strategy is a product strategy—bundle and restrict access—whereas OpenAI has been more philosophically aligned with enabling access everywhere, and a model given to 40 companies with 10,000 employees each is not private. He saw the current rationing of very large models as temporary because much larger compute clusters are coming online over the next three to five years; he theorized that Google has an Ultra model in the basement and distills from it for Flash and Pro, and he noted that the idea models would cap out at two trillion parameters is proving wrong even though the paradigm needs another scaling law. At 44:02 he called context length the slowest scaling factor in LLMs—roughly 4,000 to a million tokens over three years, with Gemini at a million for two years and few people using it—and said memory is likely the biggest constraint.

At 44:34, Swyx said he had been bearish on open models a year earlier, after Ankur Goyal of Braintrust said open-source market share was 5% and going down. He has changed his mind: the number is going up. The capability gap is hard to judge because public benchmarks are easy to game, and OpenRouter stats need price adjustment because many open models are heavily discounted. The important cohort is the top 20% of builders, which acts very differently from the bottom 80% and is moving toward open models; Fireworks, Together, and finetuning-as-a-service benefit.

At 47:41, Swyx said people are not appreciating dark factories enough. The first level of AI coding psychosis, which he saw at Cognition five months earlier, was zero human-written code; the next frontier is zero human review. Very few people are doing that, but OpenAI is exploring it, and Swyx thinks it is the only scalable way. It forces a flipped SDLC with more testing and automated verification, and it produces so much disposable software that quantity can drive quality. The people who do well in 2026 will not be the cynics who dismiss it as slop; they will bend it the right way.

At 49:47, Jacob said he had changed his mind on post-training for quality. If the single best thing an application company can do for a customer in a three-month window is a post-training run that improves the model's output, it can be worth doing even if the results are thrown out when the frontier models catch up—because the raw data is not thrown out. Swyx pointed to synthetic rubrics, Dr. GRPO, and multi-turn RL with hundreds of turns as signs that RL is becoming more domain-specific. At 51:49 he named memory and personalization, plus world models, as the next frontier; he pointed to Fei-Fei Li's essay on spatial intelligence as the right problem statement even if she does not yet have the solution. He is skeptical of equating world models with robotics or gaming; the deeper question is whether AI understands what a table, matter, and physics are. He compared it to Good Will Hunting: an LLM can know everything from books but has not lived it.