Manifesto

Dec 19th, 2024

The Opportunity of Agents

2025 will be a breakthrough year for AI agents as they reach true production readiness. Why so much excitement? Because agents offer a transformative opportunity, even greater than what SaaS once did. If agents perform as expected, we’ll move from a world where humans handle 90% of tasks and software 10%, to one where software handles 90%, and humans focus on the remaining 10%.

For enterprises, this shift is game-changing. By offloading repetitive tasks to agents, companies can free their human workforce to focus on creative, strategic, high-value work—leading to leaner teams, cutting through the noise and making better decisions.

Agents as Automation 2.0

This idea isn’t entirely new. For a decade, tools like RPA and API-based automation software promised the "fully automated enterprise." However, the reality fell short. Despite all these years, traditional automation has only addressed a small fraction of companies' workloads. Why? Traditional automation relies on rigid, predefined workflows. Each step is hard-coded, so any change or decision point disrupts the system. A new field, a slightly different form, or an unexpected exception—and the entire workflow fails.

Agents, on the other hand, reason dynamically. Instead of following a fixed process, they pursue a goal, plan steps, adapt to changes, and make decisions along the way. This flexibility allows agents to handle workflows that rigid automation tools can’t, from complex support requests to intricate data-processing tasks. Traditional automation is also difficult to build, costly to maintain, and only addresses a small portion of a company’s workflows. Agents promise to change that.

What’s Missing in the Agentic Stack

If agents are so powerful, why aren’t they everywhere? We see early examples like AI SDRs, recruiters, and analysts—but they fail when pushed beyond their limits.
To build a truly effective agent, you need three key ingredients: reasoning, logic, and tools. Reasoning comes from frontier AI models, logic from in-prompt instructions and context. But what about the tools?

Imagine a customer support AI agent tasked with answering user requests. You provide logic, context (like an FAQ), and reasoning from the model. But what if the user asks for an order status, which isn't in the FAQ? A human would simply open the back-office app, check the status, and respond.
Enabling this workflow requires a custom API integration. Worse, if there's no API, you'd need to build a custom RPA script. Then, if the user asks for something else—like filing a claim—you have to start from scratch for that new workflow.
The average employee uses 30 different applications, each with dozens of features. Building and maintaining integrations for every possible action becomes a nightmare.

Now, imagine an agent that can use any application—just like a human—without custom scripts or integrations. You give it a goal (“file a claim”), and it figures out how to navigate your software stack autonomously to achieve it. This is what we’ve been working on for the past year, and we believe it’s a crucial step in making agents smarter and more practical.

Universal, Autonomous Tool Use

Twin A1 is an action agent that performs tasks on any web application without the need for custom APIs or fragile RPA scripts. Instead of coding each step, you set a goal, and A1 finds its own path—exploring pages, clicking buttons, and entering information as needed, just like a human. With Twin A1, you shift from rigid automations to flexible, goal-driven workflows. Instead of building integrations for each new use case, you rely on A1’s universal capabilities. It adapts to new interfaces, and evolving a process now means simply updating your prompt. In just a few months of beta, Twin users have completed over 40,000 tasks across 9,000+ applications. By comparison, it took Zapier 10 years to build integrations for 8,000 applications.

While this technology will continue to improve—handling more complex tasks with greater autonomy—it already addresses a major pain point for companies building agentic applications. A1 enables agents to perform any task in any application with a simple prompt. You can easily scale your agents’ capabilities by just sending a new prompt to A1.

We’ve focused on making this solution as simple as possible for building, testing, refining, and deploying agents at high speed, reliability, and low cost.

A1 integrates seamlessly into your stack through our API or no-code integrations.Our infrastructure includes everything needed: browser management, geo-localized proxies, agent authentication, model training, orchestration, observability, and inference.

We continuously improve our agent framework, using human-annotated and synthetic data to enhance models, ensuring our customers always get the best performance out of the box. Since these systems are still evolving and can make mistakes, we've built a seamless training experience to help you scale your agents in production. A1 is easily correctable, and we also offer custom fine-tuning services to optimize A1 for your specific use cases.

Next Year Focus

In 2025, we’ll focus on three core areas:

1. Enhancing Core Performance: We’ll make our agent more reliable, cost-effective, and faster through improvements in data generation, model training, and infrastructure.
2. Improving the Agent Building and Deployment Experience: We’ll enhance key areas such as authentication, training, and observability features.
3. Developing New Capabilities: While browser control is a great tool for using any web-accessible application, it’s not the only solution. We’ll explore new capabilities to cover all use cases with a single platform.

Currently, A1 is available to high-volume customers, but we aim to offer a more self-serve version later this year.

In the long run, we believe breakthroughs are needed to create truly smart, autonomous agents that can tackle complex business goals with minimal supervision. Imagine a senior employee: you set an objective, and they make progress independently, using methods, tools, and processes you might not have considered—only seeking validation or feedback occasionally. This is our target.

To achieve this, we’ll focus on enhancing memory, allowing agents to run for longer periods, learn from their own experiments, and make better decisions. We’ll also work on generating high-quality synthetic data through agent self-exploration, reducing our reliance on human annotations. We’re working on these long-term challenges, so if you're interested in these topics, let's have a chat!