Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
2025 was supposed to be the year of the AI agent, right?
Not quite, recognize Google Cloud and Replit, two major players in the AI agent space and partners in the AI agent space. "ambiance coding" movement – during a recent VB Impact Series event.
Even though they are developing agent tools themselves, executives at both companies say the capabilities aren’t quite there yet.
This limited reality boils down to challenges with legacy workflows, fragmented data, and immature governance models. Additionally, businesses fundamentally misunderstand that agents are not like other technologies: they require a fundamental rethinking and redesign of workflows and processes.
When companies create agents to automate work, “most of them are toy examples,” Amjad Masad, CEO and founder of Replit, said at the event. “They’re excited, but when they start rolling it out, it doesn’t really work very well.”
Build agents based on Replit’s own errors
Reliability and integration, rather than intelligence itself, are the two main obstacles to the success of AI agents, Masad noted. Agents frequently fail when running for long periods of time, accumulate errors, or lack access to clean, well-structured data.
The problem with enterprise data is that it is complicated (it is structured, unstructured and stored everywhere) and exploring it is a challenge. On top of that, there are many unwritten things people do that are difficult to code into agents, Masad said.
“The idea that companies are just going to bring in agents and agents will replace workers or automatically automate workflows is simply not the case today,” he said. “The tools are not there.”
Beyond agents, there are computing tools that can take over a user’s workspace for basic tasks such as web browsing. But these are still in their infancy and can be buggy, unreliable and even dangerous, despite the accelerated hype.
“The problem is that computer usage patterns are really bad right now,” Masad said. “They’re expensive, they’re slow, they’re progressing, but they’re only about a year old.”
Replit learns from its own mistake earlier this year, when its AI encoder erased a company’s entire codebase in a test. Masad conceded, “The tools weren’t mature enough,” noting that the company has since isolated development from production.
Techniques like loop testing, verifiable execution, and development isolation are essential, he noted, even though they can be very resource-intensive. Replit has incorporated in-the-loop features into version 3 of its agent, and Masad said its next-generation agent can run autonomously for 200 minutes; some ran it for 20 hours.
However, he acknowledged that users had expressed frustration with latency. When they send a “heavy prompt,” they may have to wait 20 minutes or more. Ideally, they expressed a desire to be more involved in a creative loop where they can enter many prompts, work on multiple tasks at once, and adjust the design while the agent works.
“The way to solve this problem is parallelism, to create multiple agent loops and have them work on these independent features while allowing you to do the creative work at the same time,” he said.
Agents need a cultural change
Beyond the technical perspective, there is a cultural barrier: Agents operate probabilistically, but traditional businesses are structured around deterministic processes, noted Mike Clark, director of product development at Google Cloud. This creates a cultural and operational mismatch as LLMs embark on entirely new tools, orchestration frameworks and processes.
“We don’t know how to think about officers,” Clark said. “We don’t know how to determine what officers can do.”
Successful companies are driven by bottom-up processes, he noted: building no-code and low-code software and tools in the trenches is funneled down to the bigger agents. For now, successful deployments are small, carefully targeted and heavily supervised.
“If I look at 2025 and this promise that it’s going to be the year of the agents, that’s the year a lot of people spent building prototypes,” Clark said. “We are now in the middle of this large-scale phase.”
How to secure a world without pastures?
Another challenge is AI agent security, which also requires rethinking traditional processes, Clark noted.
Security perimeters have been established around everything, but that doesn’t work when officers need to be able to access many different resources to make the best decisions, Clark said.
“It really changes our security models, our baseline,” he said. “What does the least privilege mean in a world without grazing and defenseless?
Ultimately, the entire industry needs to rethink governance and businesses need to align with a threat model around agents.
Clark pointed out the disparity: “If you look at some of your governance processes, you will be very surprised to find that the origin of those processes was someone on an IBM electric typewriter typing in triplicate and handing it to three people. That’s not the world we live in today.”
