Coding Will Be Solved. The Bottleneck Is Saying What You Want

Zach Lloyd's opening claim is about shape, not models. The general form factor of the terminal, he says, is perfect for agentic work because everything is time-based. It is input of text and output of text. You log what you are doing. You can multitask agents in the terminal easily. That the terminal has become the center of agentic development is, in his telling, a great stroke of luck and a huge opportunity for Warp.

Sonya Huang, hosting Sequoia's *Training Data*, frames the rest of the hour around that luck. Coding interfaces are converging into a workbench built for prompting and agent orchestration. The next frontier, as she sets it up, is not developers typing prompts but ambient agents in the background, answering system events such as server crashes or security incidents. The coding market is brutally competitive, model providers are racing into the application layer, and Lloyd's larger thesis is that coding is nearly solved. The ultimate bottleneck, in that thesis, is whether humans can clearly express intent.

Lloyd is CEO and founder of Warp, a five-year-old developer startup whose goal has always been to help professional developers ship better software more quickly. The product began as a modern reimagination of the terminal. Today he will call it a terminal with agents built in, or, more generally, a workbench for building software with agents. In a prior life he was a principal engineer at Google and ran engineering on Google Docs. He has always used the terminal. He does not claim to be good at it.

He worked with people who were, and he saw how much a developer can get done because of where the terminal sits in the stack. Used well, it is super powerful. The classic version, in his view, is a horrible product: hard to learn, easy to make mistakes in, and the mouse does not work. He wanted to build something impactful for developers, something that helps more good software exist, and reimagining the terminal felt like a cool thing to take on.

The original business-model thesis was multiplayer. Coming from Google Docs, he reached for a Postman-like analogy: a collaborative platform around the terminal, where teams share commands, runbooks, and incident-response manuals. Warp shipped that, and it turned out to be useful not only for people but for agents, as baked-in knowledge. Where they actually started, though, was the hands-on-keys interaction. The first year to a year and a half went to input, output, and making the tool easier without diminishing its power, all of it before generative AI and coding models took off.

Huang asks whether coding models change how important the terminal is as a workbench. Ironically more important, Lloyd says. It has become the preferred form factor for working with agents. You can work with them in the IDE or in the terminal, or you can treat Warp as a broader workbench that started as a terminal. The reasons are the ones he opened with: time, text, logs, and easy multitasking. For a company that spent years on a pre-AI bet, the shift is luck.

The old distinction between terminals and IDEs is morphing. Warp grew from the terminal and added IDE features, a code editor, code review, a file tree, and gets yelled at on Twitter for the file tree because it is not a pure terminal thing. Cursor, which started as an IDE, now looks a lot more like Warp in his reading: the primary interface is chat, talking to your computer, with file editing still there. He does not predict that the terminal dies or that the IDE dies. What he feels strongly about is that the form factor is changing to match the agentic workflow: prompting, adding context, reviewing agent-generated diffs. Team features matter more as agents are launched not only locally but by system events. The workbench will look more like a terminal than an IDE, he thinks, and will not strictly be a traditional version of either.

Huang's mental model is that the terminal is the chatbot, chatting with a coding agent and handing off tasks, and the IDE is a GUI for actually editing code. Lloyd says that is definitely where things started. The IDE is Microsoft Word for your code; the terminal is chatting with your computer. For professional agentic development, which he distinguishes from vibe coding, you still want both. The industry is not yet at a spot where you can be so disconnected from the code that you do not need a way of falling into hand editing. Hand editing has become a fallback or secondary interface. The primary interface is now the prompting interface. Product-wise, it is all merging.

Why stay on pro developers? He cares about helping build the software he uses every day, the ten or so apps in his Mac dock and pinned Chrome tabs: Google Docs, Spotify, Notion, Figma, Warp. Those hard-to-build apps, he says, are what the world spends most of its time in, and they are built more by professionals and more by enterprises. Vibe-coding tools such as Lovable or Replit are empowering, anyone can make an app, but the economic value is lower. He does not think he has ever spent a day using an app built in a no-code, low-code, or vibe-code tool, while he lives in software built by pros for immersive, hard, economically valuable use cases.

Huang calls the coding market the most brutally competitive subset of software she has ever seen. Lloyd orients Warp as a general-purpose agentic development workbench: you can use it like Cursor, you can use it like Claude Code. The unique claim is origin. He says they are truly the only platform that grew out of the terminal. A lot of products grew out of the IDE by forking VS Code and look very similar. A lot of others are text-based apps that run within the terminal, and those are basically all the same. Warp's approach shines, he says, on traditionally terminal-heavy work beyond coding: setup, deployment, Docker, Kubernetes, incident response, backend DevOps, SRE, production. At any given moment they are one of the top five agents on SWE-bench and typically number one or two on Terminal-Bench. They are trying to compete on product quality. There is also competitive pressure around cost, which is really challenging.

How do you compete when Anthropic can subsidize a tool with model profits? It is Anthropic, OpenAI, and Google, he says. One answer is to sit in a more premium part of the market. Coding agents and developer experience are not a commodity; they are not totally fungible; the product experience matters. Another is to stay away from the most cost-conscious segments, vibe coders running agents 24 hours a day on prototypes, which is not a professional developer's usage pattern. For a pro, 20, 40, or 80 dollars more is tiny next to the productivity. A third is to sit above the model providers. For three to six months Anthropic was the main show in frontier coding; now Gemini 3 and the latest Codex models are, in his view, basically on par with the latest Claude. Letting people choose, and routing to cheaper open-source models, is an advantage. If you view it as a cost race among identical tools, Warp has to differentiate harder on product and orchestration. He does not think that is quite the situation.

They have about 700,000 active developers, a funnel from the terminal into coding use cases. Terminal-Bench is a harness on a mix of models, plus the capabilities of the app. The bench is not just coding; it is all sorts of terminal work, and Warp has an intrinsic advantage by *being* the terminal rather than an app inside one. One task was playing Zork, an interactive terminal game. They can do computer use at the layer of the terminal, the way browser-use companies do it at the layer of a web page, which is hard for other harnesses. Then came the pricing redo. The old plan was a subscription with a fixed pile of AI credits, priced so that full utilization would lose money and average utilization, 20,000 of 50,000 credits, would not. People used more and more. Revenue was adding about a million every five days, later slower. They talked with Andrew at Sequoia about raising more and playing the game. Margins were bad. A cost race was not a race they could win. The new shape is a $20-a-month base plus credits on top, so more usage is healthier. It is more expensive. There were a lot of complaints, which sucks, he says, but they could not keep subsidizing. Growth is still pretty good, and now it is sustainable.

Would they train their own models? Fine-tunes and RL, yes, like some competitors; a full frontier model, hard to imagine given the capital. The strategic asset is data on terminal workflows and agent interactions. They will lean into a mixture of models and routing across latency, cost, and quality. The philosophy is a great default, then control, because these are developers. There is an efficiency default and a performance default, then raw choice. About 70 to 80 percent of users leave auto alone. Auto currently mixes GPT-5.2, in the Codex family, and Sonnet 4.5. Among people who pick, Gemini 3 Pro has been very popular lately. Stack-ranked, Anthropic is still most popular, with a decent split opting into Gemini and OpenAI. Grok is not in Warp. xAI has reached out a bunch of times. He is not opposed, but every model means tuning the harness, and he wants a concrete benefit to users. The harness is prompting language, tool definitions, context management, when to use a sub-agent with a separate window, when to summarize, when to truncate a gigantic terminal command, RAG, MCP. You make it good by measuring: internal evals, public benchmarks, and user data in Braintrust, turning failure modes into replayed evals. That data-driven shift, he says, was 100 percent necessary. Tab autocomplete is not a high priority; the more typical hand-editing use in Warp is reviewing an agent's code, not typing it.

Huang asks what other UI innovations he sees as the Word-like IDE and the chat-with-your-computer terminal merge. The biggest change in the next year, Lloyd says, is more cloud agents, ambient agents, already an investment at Warp. Rather than a developer at a keyboard giving a prompt, a system event triggers an agent: a server crashing, a cluster of user reports, a security incident filed against you. Those events become context for an agent that launches not on someone's laptop but in the cloud. That implies the workbench becomes more of an orchestration platform, a cockpit for managing not just your own agents but the team's. You will need a strong team concept. The workflow will not be sit at a desk, write a change, push a PR. Agents will push PRs, leave an initial round of reviews, and file tasks in the tracking system. All of it needs tracking, coordination, and new integrations. This evolution off interactive agents into cloud agents is probably Warp's biggest product focus for next year.

It is a massive infrastructure push, turning Warp from a product into a platform. They think in layers: an agent SDK; agent hosting, so a smaller company that does not want to stand up cloud spots can let Warp host the work, a category of startups he finds interesting because it means the thing is real; an API layer for status, takeover, progress, logs; and a management layer for who started what, what state it is in, and which PRs it produced. Person-driven agents do not go away. Ambient cloud agents will take toil and one-shotable tasks. Harder, more interesting software engineering, he thinks, will still be done by a developer at their workbench.

On the scaling-law chart that matters to Huang, how long can the agent run, Lloyd's number for real coding tasks is 20 to 30 minutes at the max that still works. You can let it run longer; it will start going in circles. Context is still limited, and unattended running is costly. By far the best results come when the agent is steered: an upfront plan, then check-ins. Hours of work only make sense to him on a really clearly defined task, a big code migration, a big job. Internally they are debating whether the cockpit is one product or two. Orchestration feels more web-centric, Warp's primary interface is not the web, and the user is sometimes different. The advantage of bundling is a seamless handoff from cloud to local. A common workflow already: tag something in Slack or Linear, ask for a button-position fix, the agent works in the cloud, a developer ties the loop on the local machine. The UI will be a little more like task management. He does not buy task management as the primary primitive. He does not think every developer will live in Linear and Jira. They will want some view of what the agents are doing across systems.

Huang asks where frontier models are good and where they still produce errors. Lloyd uses the stuff constantly, biased because Warp's codebase is a very custom, large Rust tree, which he still thinks is an interesting vantage. Agents do medium-complexity tasks pretty well with a bunch of guidance. They cannot do whole big projects, at least Warp has not had success there, and he does not trust them with very fundamental architecture decisions. You want constrained tasks, but they are well beyond trivia such as changing a button color or some text. They can make apps. They are very good at zero-to-one. They can solve kind of hard bugs. He tagged an agent in Slack to add a new slash command to Warp; it produced a 300-line PR that was basically right. On a scale of 0 to 10, he puts the field at a six, maybe. It is real and game-changing. It is not at the level of what a full-time engineer on a hard product needs to do.

Context window is still a big issue, and even with bigger windows, reasonable attention over the whole window is hard. There is no continuous learning. Memory is a slow, inefficient repopulation with a bunch of files. Each run is a big stateless thing that you fill up before you set it loose, which stinks. How to use the tools is still early; the stuff did not exist a year ago. Context engineering, how you set up projects so agents can work with them, is a problem. Across Warp's own team, usage is high variance. They have rigorous standards for writing code and almost no standards, despite trying, for how to use the agents. No one has been taught. There are not even agreed best practices. That, he says, is pretty nascent.

Huang's vibe-coding experience is still a lot of errors, and she wonders whether verification would let you RL them away. Lloyd says errors are definitely still there. An interesting milestone: it is pretty infrequent now that the agent produces something that does not compile. Four or five months ago, getting to a compiling version was a problem. It compiles for him about 100 percent of the time, which is amazing. It still produces bugs relatively frequently, and it does not have a good way of closing the loop on whether the thing works. Some version of browser use or computer use, where the agent verifies the change from the user's perspective, not the code perspective, is pretty important. Model providers have beta computer-use APIs; Warp is looking at it. It becomes even more important as more work is done remotely, because the real pain of remote work is verifying from a user perspective. With that loop, RL toward behaviorally correct, not just statically compile-correct, gets easier.

Will there be a superintelligence moment where models are better than the best human coders? He has no idea. What he does think is that coding will be solved by models. The limiting factor will be expression of intent from humans: what do you want built, how do you build it, how do you say that clearly. English is ambiguous. Huang objects: is coding not the truest expression of intent? Yes, he says, but the world is moving from speaking in code to speaking in English to build apps, reintroducing ambiguity because people no longer express what they want directly. They go through a translation layer. It is a step backwards in a sense, and way more efficient. He thinks you will get to a point where you do not need to be on the frontier to match a person's intent as well as possible, which would be a bad place to be in the API business for coding tokens: you will not charge a huge margin. That is why Anthropic, OpenAI, and Google are going so hard at the application layer. Huge risk at the API layer, for this vertical, which he thinks is basically solved within a few years. He is prognosticating, he adds, not asserting. Enterprises still treat the spend as a productivity boost, evaluated with engineer feelings or DORA metrics, not as labor. Pitch a $200,000 agent to replace a $200,000 engineer today and they would say not even close. What would change it is products launched with very minimal engineering, a proof point he does not want, because he is an engineer at heart and does not want people losing their jobs.

Huang closes on coding as an art form and a 2023 blog post she liked, on the future of productivity interfaces as ask and adjust. Lloyd wrote it shortly after ChatGPT came out, as Warp started integrating the model deeply. The idea, obvious now, was that productivity interfaces had always been geared toward hand editing: drawing vectors in Figma, entering cells in Google Sheets, typing code in VS Code. The thesis was that the primary interface would become AI-based. You ask the app to do the thing. You, the human author, adjust. Adjusting might mean re-prompting, or, if prompting failed, dropping into the old hand-editing interface to complete the change. He did not have the word agentic at the time. For coding, he thinks that is where we are: you start by asking, then you adjust.

He also asked whether you could get rid of the adjustment piece. The places you would need it least, he thought, were domains with many acceptable solutions. Ask for an image and a thousand outcomes might work, so you re-prompt until you like one. Code or a spreadsheet, where one thing needs to be right, would keep the ability to get it perfect by hand. Three years on, he thinks the thesis was not bad. It has held up okay. Warp did coin something: agent mode. They were the first product to launch a branded feature with that name. Ask ChatGPT where it came from, he says, and it came from Warp. It is now a very common way of describing the feature. He wishes they had trademarked it and were getting kickbacks. Huang thanks him for the observations on a white-hot market, and for the case that the terminal will be the workbench.