Anthropic's Vision for AI Agents and Platforms

This API was designed for long-term abstraction, handling complex and evolving system interactions.

She observed a customer integrating the API using Cursor, an AI-powered code editor. The automated integration achieved approximately 70% accuracy, a significant feat at the time. This demonstration was a pivotal moment for Jiang, convincing her that the future of complex coding tasks and knowledge work lay with AI.

"And it came back and it was like 70% accurate. It was like a you know a while ago and that was like just eye opening for me. I was like the future is here and this kind of like knowledge work this kind of like just capabilities being unlocked by the models like I want to be part of that."

The company aims to tackle undifferentiated infrastructure challenges, allowing developers to allocate their efforts to unique aspects of their products.

Building infrastructure for long-running agents presents a significant distributed systems problem, often requiring extensive resources. Anthropic views this foundational layer as undifferentiated work, believing it should not be the primary burden for developers.

Instead, developers should be empowered to concentrate on core business logic and creating exceptional customer experiences. Angela Jiang emphasized this, stating, "you should do as a builder on top of the platform, you should do the things that differentiate you and you shouldn't have to do the things that are undifferiated."

Traditional execution sandboxes are typically ephemeral, making them unsuitable for supporting long-running, complex tasks without constant re-establishment and potential for failure.

Anthropic's platform design introduces a modular architecture that separates the agent's 'brain' from its execution environment. The 'brain' runs on durable, persistent servers, ensuring continuity and state retention. When the agent needs to perform an action, a transient sandbox is spawned for execution.

This design ensures that if an execution sandbox fails, it can be torn down and replaced without compromising the entire agent's state or progress. Katelyn Lesse explained, "what we've spent a lot of time and energy on is how do we rethink the agent make it a more kind of modular broken down architecture and say maybe like the harness or the brain of the agent runs over here on a server that's durable and then when it needs to execute work which is a scary part go spawn a sandbox do it within the sandbox and then tear down the sandbox when it's done"

Managed agents prevent premature complexity for early-stage users by handling much of the underlying technical detail. As users' needs evolve and they require more granular control or customization, the platform allows access to lower-level components and configurations, offering a flexible and scalable path for growth. Angela Jiang clarified: "Yeah, we would actually just tell you to use cloud manage agents. Um, like literally this is actually one of the reasons why we designed this product. It's a set of higher order abstractions and you can access the lower layers."

The platform team focuses on abstracting technical details, making it easier for developers to build and deploy long-running agentic applications.

This harness reduces the need for manual prompt caching and context management, which are common pain points in agent development. Users can tune system prompts, define specific skills for their agents, and utilize Multi-Cloud Provider (MCP) connections for seamless external data integration, all facilitated by the managed agent framework. Angela Jiang stated that "the point of cloud manage agents it's like a harness that like pretty generically is good at that stuff and like holding on to a model and letting it do longunning work."

Anthropic clearly defines its ownership boundaries within the agent ecosystem. The company maintains internal control over core model serving and safety classifiers, recognizing these as critical components directly related to its core product and responsible AI principles.

However, Anthropic encourages users to self-host elements like sandboxes and MCP tunnels. Katelyn Lesse explained that above this foundational layer, Anthropic becomes opinionated about architecture but not necessarily the underlying infrastructure. The goal is to provide robust architectural guidance while offering flexibility for users to manage their own infrastructure, fostering a balance between guidance and autonomy. "Everything above that. I think that's where we're in the territory of being opinionated about architecture but not necessarily opinionated about infrastructure."

Anthropic explicitly addresses data privacy concerns, particularly regarding the use of closed models. The company firmly states that it does not train its AI models on user data.

This direct stance aims to counter the misconception that using closed models could lead to a loss of corporate advantage due to data exposure. Angela Jiang directly stated: "we don't train on your data. So, it's like just straight up we we do not do that."

The argument is that while more advanced models might appear more expensive per token, they can be more cost-effective if they reduce errors and achieve desired outcomes more reliably.

The business objective should be overall efficiency, not merely minimizing token counts. High-intelligence models can prevent wasted tokens by completing tasks correctly on the first attempt, thereby reducing iterative corrections and retries.

Angela Jiang stated: "we think that probably more likely is that we need to get a better job at understanding the outcome that you want and then from that outcome make sure we take the absolute most efficient path." The long-term goal is to automatically identify the optimal model and path for specific tasks, further optimizing resource usage.

This approach, often called an 'advisor-executor' strategy, optimizes token usage by leveraging the strengths of various models.

For example, pairing a smaller, more cost-effective model like Sonnet (as an executor) with a larger, more intelligent model like Opus (as an advisor) can achieve near-Opus level performance. The advisor model guides the executor, teaching it to perform tasks more efficiently, resulting in reduced token consumption.

Angela Jiang provided an example: "we've got some evals that show like sonnet executing with opus advising ends up getting almost opus level performance and it's actually cheaper than just sonnet because opus taught it how to do its job better and it used less tokens to get the job done." This combined approach can be cheaper than using a single large model via brute force, as the advising and grading roles reduce overall token consumption.

At its most basic, a harness is a loop that facilitates continuous interaction and task execution. Katelyn Lesse jokingly referred to it as "a loop," but clarified that "a harness really is like the most basic version of a harness is like a while loop that's literally just like go back and forth between like get input from the user, ask the model what it thinks, then call a tool over here, and you keep that thing running."

Beyond this fundamental loop, advanced harness requirements include ensuring environment safety, maintaining persistent conversation states across interactions, and securely handling credentials for tool access. These components collectively allow agents to perform complex, multi-step operations reliably over extended periods.

A common pitfall in AI agent implementation is attempting to mimic inefficient human processes. Many organizations try to "take a very human process that's been duct taped together and it it works and it has all these policies uh and then you just try to like insert agents exactly where you see human inefficiencies," Angela Jiang observed. This approach often leads to agents inheriting the same inefficiencies and limitations.

Instead, Anthropic advocates for an 'agent-first' design, urging developers to break down ambitious projects into basic steps and redesign the entire process from scratch. This allows for a more optimal integration of AI capabilities, rather than simply automating existing, often convoluted, workflows.

Successful agents are designed to act autonomously, only escalating to human intervention when absolutely necessary. This maximizes their efficiency and prevents them from becoming bottlenecks within pre-existing, suboptimal systems.

Angela Jiang articulated this vision: "the kind of through line that we sort of see is you almost want it to be uh almost like slightly more human ironically or maybe not ironically since it's uh it's AI it's artificial intelligence and you wanted to get to the place where it acts a little bit more like a really amazing co-orker."

This shift moves away from ubiquitous chat boxes towards more direct and proactive agent engagement. The ideal agent acts like a proactive team member, taking initiative, handling tasks efficiently, and providing context when needed. Platforms like Slack are cited as environments where humans naturally collaborate, making them viable spaces for agents to integrate seamlessly and act as intelligent collaborators.

Katelyn Lesse highlighted that "there's also just going to be this like continuation of agents can do stuff for you when you're not paying attention to them. And I think in that world, we're just going to continue to need to have ways to say like, but how do I like set off an agent to go do a thing that like is going to come back with what I actually care about, right?"

Agents are increasingly capable of operating in the background without constant human oversight, freeing users from micro-management. However, the core challenge lies in designing intuitive ways for users to accurately set up agents to deliver precisely the desired results. Finding the optimal user interface and interaction patterns for managing these background operations is an open problem that Anthropic continues to explore.

Angela Jiang described this as "the most magical outcome that is probably like most true to growing model capabilities is that the only thing really helping you is actually just the model. So, in the most pure sense, the way that you would solve any problem is you would just literally talk to Claude and Quad would do everything."

In this ideal state, models would not only help users express their needs but also assist in building the necessary scaffolding and even constructing entire products. Currently, product deployment and human-deployed engineers serve as transitional solutions, bridging the gap to this future.

This vision suggests a paradigm shift where the AI model itself becomes the primary tool, capable of understanding, planning, and executing complex tasks directly based on user input, simplifying the entire development and problem-solving pipeline.

For common business challenges, productized solutions offer quick wins and established efficiencies. However, for complex, foundational problems, custom AI-driven processes become indispensable.

Angela Jiang highlighted scenarios like massive code migrations, which could "take literally like 12 years uh for us to go and accomplish and it's a pain... and as the end result will be if I can solve this like code migration problem, I really have also unlocked the company to do something they never could believe was possible."

In such cases, custom AI engineering can enable radical speedups, transforming years of work into months. This strategic investment in AI allows companies to tackle previously impossible challenges, unlocking new capabilities and competitive advantages that off-the-shelf products cannot provide.

Katelyn Lesse explained, "And now it's like you kick off that project, you have a sense for where you want to go and engineers are kind of like, okay, cool. It's done like two days later and you haven't had the time to do all the work to get everybody on the same page. And so that I think is where we've definitely been on a journey within our team to figure out how do we just solve for that?"

Traditional project workflows, where coordination activities (by Product Managers, Tech Leads, etc.) often run in parallel with development, are being disrupted. The speed at which AI-powered engineering teams can execute tasks means that the coordination and alignment phases can now become the limiting factor.

Teams are grappling with how to realign roles and processes to match this newfound engineering velocity. This involves rethinking how decisions are made, how stakeholders are brought into agreement, and how to prevent the rapid pace of development from outstripping the organization's ability to coordinate effectively.