pre.dev @predotdev

@parzival1213 Hey! Nice project, will check it out. Here is our benchmarking suite. Would be interesting to know if you think we add something. We are trying to cover most real-life use cases. github.com/predotdev/brow…

Look at this: Browser Agents running right inside your AI development project. One of the most powerful capabilities within predev is the ability to deploy native browser agents. Your primary coding agent can spawn multiple, parallel browser sub-agents to execute web-based tasks without ever breaking your current development workflow. Here are the four most popular engineering use cases our users are leveraging right now: 1. Autonomous Documentation Research The most immediate use case: sending agents to research deep technical documentation, API references, and open-source libraries specific to your tech stack, ensuring the code written matches the latest upstream versions. 2. Visual E2E and Regression Testing True continuous integration. As your application is being built in the cloud sandbox, browser agents can spin up a headless instance, take visual screenshots, verify the layout looks flawless, and click through critical user paths to ensure everything works exactly as specified. 3. Schema Validation & Seed Data Ingestion Instead of manually writing tedious mock datasets, browser agents can ingest public reference data, synthesize complex datasets, and automatically populate your database tables while strictly validating the data against your exact backend schemas and boundaries. 4. Extracting Design Tokens & UX Benchmarking Instead of building design systems from scratch, you can point your browser agents at industry-standard design web references. The agent analyzes abstract layout structures, extracts clean design tokens (like color palettes, typography scales, and component hierarchies), and maps out optimal user flows to help you architect a completely unique user experience for your own project. The best part? Just like our coding agents, these browser sub-agents run entirely in isolated cloud environments. You never have to grant a local agent access to your physical machine, risk leaking your local IP, or drain your local memory. You can close your laptop, walk away, and check back in when the parallel execution is complete. We built predev's native browser agents from scratch. By optimizing the orchestration harness layer to stream compacted DOM states rather than raw, heavy page layers, we’ve achieved exceptional performance at a fraction of standard token costs. (More technical details on the harness efficiency in the comments below). Head over to predev to experience browser-driven cloud development. Let us know what you are building next.

x.com/predotdev/stat…

pre.dev @predotdev

2 months ago

At predev a lot of our harness engineering effort goes into maximizing token runway. But is the juice worth the squeeze? 👇

1 0 1 162 0

Use the predev native Kanban board as the top-level control plane for your coding agents. When you are planning large software projects to build with coding agents, Behavior-Driven Development and user stories are the gold standard for achieving predictable results. This holds true whether you are building with a team of human engineers or a flock of agents. Everyone loves jumping straight into writing code. But if you want to ship production-grade software without endless refactoring, starting at the planning layer is the winning strategy. What it consists of: A structured framework detailing the persona (who wants it), the action (what they want to do), the value (why they need it), and a deterministic checklist of binary acceptance criteria. An Example: As a returning shopper on the checkout page, I want to apply an alphanumeric promo code to my cart, So that I can see my order total update before finalizing the purchase. By structuring tasks this way, the focus shifts to Verification over Implementation. You are managing by outcomes. Don't worry, the predev Architect Agent writes these user stories for you based on your abstract ideas. However, understanding this structure is vital because it unlocks two leverage points: 1. A Unified Language for Humans and Agents LLMs are exceptionally good at semantic comprehension. By feeding them user stories, you give them room to define the optimal how. This often results in elegant, novel approaches you might not have considered. More importantly, it gives the agent a clear, binary checklist to automatically verify its own code before pushing a PR. 2. True Task Idempotency Because well-scoped user stories exist independently of the volatile state of the codebase, they remain semantically stable. The agent can evaluate a single card on the Kanban board within the broader context of the system architecture without getting tripped up by concurrent changes elsewhere. This is vital for preventing agent drift. This clean isolation allows you to completely parallelize your roadmap. You can tear individual user stories off your Kanban board and assign them to dedicated, isolated coding agents simultaneously. They work in their own cloud sandboxes and merge their completed features back via individual pull requests. This gives you granular control, code attribution, and a bulletproof continuous integration pipeline. This underlying board framework acts as the shared memory that enables long-horizon agents to execute and verify their own work autonomously. Beyond the Kanban board, predev provides additional levers to steer your agents at scale, including defining your core technology stack, modifying your global product specifications, fine-tuning agent execution profiles through customizable "skills," and spawning sub-agents for multi-session builds. Head over to @predotdev to let the Architect Agent generate your Kanban board, and start steering your coding agents from your control plane.

@adampredev We open-sourced all terminal-bench trajectories on GitHub. Details in the post below. x.com/predotdev/stat…

pre.dev @predotdev

2 months ago

If you have been using Claude Code professionally, take a minute to read this. We beat Opus with Sonnet by using the predev harness. Here is what it means for agentic coding: Orchestration beats brute reasoning. A smaller model running on our architecture just beat Claude Opus

2 1 3 1K 0

x.com/predotdev/stat…

pre.dev @predotdev

2 months ago

If you have been using Claude Code professionally, take a minute to read this. We beat Opus with Sonnet by using the predev harness. Here is what it means for agentic coding: Orchestration beats brute reasoning. A smaller model running on our architecture just beat Claude Opus

2 1 3 1K 0

And btw. We also beat Claude Code + Sonnet with our own harness Haiku. Sharing our full breakdown of the results and all open source trajectories in the comment.

Adam Elkassas @adampredev

3 days ago

In light of recent discussions around open-weights vs. frontier labs, the pendulum has swung completely. The risk posture has changed. Organizations used to worry about geopolitical AI risks; now, they fear the labs themselves. The worry is that labs will verticalize into

1 0 2 155 0

One of predev's most powerful features is the Architect Agent. It takes a raw product idea or briefing and decompresses it into a complete product roadmap, architecture graph, a recommended technology stack, and a comprehensive specification. While the Architect Agent does the heavy lifting, we want to share a few helpful tips to get you thinking about your next project while dramatically improving how you instruct AI agents on what you want. An effective prompt should answer five fundamental questions. First, who is it for? This defines your exact target audience and user personas. Second, what does it solve? This specifies the core friction or pain point your application addresses. Third, why does it matter? This establishes the ultimate value proposition or business impact. Fourth, what is it? This outlines the primary product type and its core functionality. Fifth, what is it not? This draws a strict boundary around your project scope to prevent feature creep from day one. Let us look at a simple example to visualize what an effective, high-input prompt looks like. Imagine prompting: Create a web application for content creators called PostSnap. It takes a raw LinkedIn post URL, fetches the text and author profile, and instantly renders it into a beautifully styled, high-resolution downloadable image with customizable gradients, shadows, and clean background framing. Instead of settling for messy, pixelated manual screenshots, creators use these polished graphics to instantly cross-promote their written insights across visual platforms like Instagram, Threads, or X without losing asset quality. It is not an analytics tool or a scheduling platform; it is strictly an aesthetic image wrapper. This brief description gives you a clear mental model of the product, right? Well, the AI operates exactly the same way. Better inputs get you better outputs. And do not worry, if anything remains unclear, the predev Architect Agent will ask you targeted follow-up questions to fill in the technical blanks. Once the agent has fully decompressed your idea into a product roadmap, the tasks, user stories, and acceptance criteria are automatically piped into the predev coding agent to actually build the project. But we will cover that execution workflow in a separate post. For now, if you want to see what a professional roadmap looks like for that product idea you have been sitting on, head over to predev and try out this prompting framework. Happy building. PS: If you prefer to stress-test your ideas inside your own existing development environment, simply grab the @predotdev Architect MCP server. You can plug it directly into your existing AI tools to automatically generate technical specs and Kanban tasks before you start writing any new code

Who here has solved long term memory? No takers? Ok, we solved it - ask us how?!

AI adoption is a given, but skyrocketing token costs aren’t. Learn how high-agency engineering teams use infrastructure and multiplayer culture to drive true token efficiency. By @ArjunRajJain

Arjun Raj Jain @ArjunRajJain

a week ago

x.com/i/article/2072…

0 0 1 278 1

60+ hour coding agent on autopilot, no human interaction needed. Head over to predev to see long-horizon coding agents in action.

Arjun Raj Jain @ArjunRajJain

a week ago

Exciting milestone: Last month, we had a customer run predev on autopilot for 60 plus hours. While this scenario might sound extreme, it validates our core thesis: true, long-horizon coding agents can and will build entire software systems with zero human intervention. To

0 0 1 227 0

@grok briefly describe Axelrod's Game Theory Tournaments

Idea: Instead of benchmarks we should have models compete against each other in game theory tournaments. wdyt?

Fine-grained agent steering happens before the first line of code is written. It happens at the planning layer. Building software with predev is engineered around two distinct phases that give you total control over the coding agent's execution. The first phase is led by our Architect Agent. By describing your product concept and refining it through targeted scoping questions, the platform generates a comprehensive product roadmap. This includes granular user stories, database schemas, and technical specifications, all automatically organized into an interactive, native Kanban board. You can review every task individually, altering the scope, fine-tuning the technical constraints, or modifying the specific acceptance criteria. Once you are satisfied with the directional plan, you transition to the second phase: launching the coding agents to execute against the board. Because the requirements and guardrails are perfectly mapped upfront, the agents build with maximum precision and minimum token waste. Bonus tip: predev supports Multi-Session Builds natively. This means you do not have to wait for a single agent to work through your backlog sequentially. You can spin up multiple parallel sessions, assigning individual user stories to dedicated agents simultaneously. Head over to predev to use the Architect Agent for your next build.

low effort on Claude Code >>

Adam Elkassas @adampredev

a week ago

sometimes i feel like a @claudeai cheerleader

1 0 2 135 0

I’ll be bahhk

Adam Elkassas @adampredev

a week ago

Fable 5

0 0 0 96 1

@adampredev reducing time to market feels 1) more relevant than ever 2) more achievable than ever

One workflow that is surprisingly neglected when working with coding agents is strict task isolation, specifically, confining agent tasks to individual branches and merging them back via pull requests. The benefits of this approach are obvious to any engineer. A pull request forces the agent to focus on a narrow, well-defined set of instructions, whether it is a discrete feature request, a bug fix, or a specific user story. This makes the code highly reviewable, perfectly attributable, and easier to verify. It allows your existing continuous integration scripts to test code quality, styling, and functionality before anything hits production, catching second-order bugs and mitigating codebase pollution before it can compound into technical debt. But there is an even bigger architectural reason to enforce this workflow when scaling coding agents. We call it Multi-Session Builds. Because predev coding agents execute in the cloud, they are never bottlenecked by local machine hardware or client instances. Each agent gets a dedicated, secure cloud sandbox provisioned with the exact infrastructure resources required for the task at hand. This unlocks a powerful paradigm shift: multiple specialized agents, operating with a shared understanding of the overarching product, can simultaneously create their own branches, execute tasks in parallel within their isolated sandboxes, and merge results via pull requests. Instead of relying on a slow, sequential, monolithic approach, where context drift is inevitable as messages pile up, you can now tackle your product roadmap concurrently. This distributed execution increases output velocity while ensuring that every line of generated code remains reviewable. Head over to the predev blog for a deep dive into how Multi-Session Builds work under the hood and how you can deploy them for your next project.

@tacurran Absolutely! And we have some more exciting releases line up.

Early on with predev, we made a big bet. We believed the absolute best environment for coding agents would be cloud sandboxes. Today, that bet is starting to pay off. Here is the reasoning behind why we chose a cloud-native architecture. Zero Tooling Setup Unlike traditional coding agents that run locally through a CLI, we were convinced that for agentic coding to achieve mass adoption, the onboarding friction had to be zero. No local dependencies, no wrestling with environments, and no individual user configurations. By delivering a coding agent through a browser-native interface, it works out of the box with batteries included. Engineering teams get a reliable, opinionated workflow from day one without lengthy skill setup or agent configuration before the first line of code is written. Parallelization & Multi-Sessions Our engineering philosophy evolved straight out of our planning layer. Because predev takes a high-level concept and fans it out into a full product roadmap, complete with system architecture, specs, user stories, and acceptance criteria, one thing became clear: coding agents will collaborate just like human engineering teams. Complex builds require narrow tasks allocated to dedicated resources in parallel, not a linear, sequential progression handled by a single, mythical all-rounder agent. This made cloud sandboxes an necessity. Local machines, no matter how beefy, cannot sustain multiple simultaneous agent sessions with their individual context, tools, and heavy memory footprints. The cloud allows us to spin up isolated sandboxes with dedicated, efficient resources, handing over concise subtasks to agents all at once. Long-Horizon Autopilot Because our planning layer maps out the entire product architecture upfront, our system can reliably predict the coding agent's actions ahead of time. We turn this into an execution graph. When you pair this predictive roadmap with recursive memory compaction and tool optimization, you get true long-horizon agents. We recently had a user run predev for over 60 hours straight. A local terminal is simply the wrong environment for that scale of execution. The likelihood of a machine disturbance over a three-day period is incredibly high, forcing users to constantly babysit the terminal. Moving the coding agent to the cloud decouples the workload from your physical hardware. You can close your laptop, walk away, and check in on the agent's progress from your phone at any time. If you want to explore the capabilities of cloud coding agents, check out what we are building at @predotdev.

Context: x.com/theo/status/20…

Theo - t3.gg @theo

2 weeks ago

I’d estimate we’re ~6 months from most devs moving their code agents off of their laptops

616 131 6K 2.2M 1K