SIAgentNews.com
← All news
ConfirmedAgents

How AI agents are reshaping the developer workflow at Block

Angie Jones explains how junior developers are using agents to solve complex code problems while senior engineers focus on oversight and standards.

· First reported by The New Stack

10new hires at Block in the experiment

The short version

  • New hires at Block used AI agents to deliver solutions on their first day of work.
  • The Model Context Protocol (MCP) provides a standard for connecting agents to various applications.
  • Experienced engineers are shifting from writing every line of code to reviewing agent-generated patches.
  • The Goose project uses an issue-first process to ensure human maintainers agree on designs before agents generate code.

Why it matters

This shift suggests that the value of a developer is moving from manual coding to 'agent orchestration.' For the average worker, it means AI will increasingly handle the 'how' of a task while humans focus on the 'what' and the 'why.'

A new way to onboard talent

A lead developer at Block recently conducted an experiment that changed the traditional onboarding process for new software engineers. They gave 10 new hires a codebase they had never seen before and provided them with access to AI agents. The goal was to see if they could solve specific problems brought to the company by customers in person.

The results were immediate, as all 10 new hires delivered solutions within a day and a half of starting their first day. This scenario was previously considered impossible for people who had no prior experience with the company's specific internal systems. It highlights a shift where the barrier to entry for complex tasks is being lowered by autonomous tools.

This experiment raises significant questions about the future role of junior developers in the tech industry. If a new hire can use an agent to navigate a massive codebase instantly, the focus of their training may shift. Instead of learning where every file is located, they may focus on how to effectively direct an agent to find the right answers.

Standards for agent coordination

Angie Jones, the vice president of the Agentic AI Foundation, distinguishes between simply using an agent and building one that others can rely on. She emphasizes the importance of standards like the Model Context Protocol, or MCP. This protocol allows developers to connect agents to different applications in a consistent way.

By using standards like MCP and Agent2Agent (A2A), companies can create interfaces that remain functional even as underlying models evolve. Jones notes that these standards allow developers to say they can build on top of a system now because the foundation is stable. This prevents developers from having to rewrite their entire infrastructure every time a new model is released.

However, building reliable agents still requires significant engineering effort to manage the logic of the tool. A developer must understand the loop where an agent chooses an action, looks at the result, and decides on the next step. Engineers must decide where to place constraints on that loop and what specific permissions the agent has regarding sensitive customer data.

The shift to human oversight

At Block, Jones approached the adoption of these tools by having a small group of engineers improve the development environment for agents. By baking these improvements into the system, they made it easier for any developer to point an agent at a repository. This allows the entire team to benefit from collective knowledge rather than every individual having to learn every new prompting trick.

The role of the experienced engineer is evolving into one of high-level review and governance. While a new hire might be able to get an agent moving quickly to generate code, a senior reviewer is needed to determine if those changes are safe or useful. The human's job becomes one of quality control and ensuring the agent stays within the desired boundaries of the project.

This dynamic creates a symbiotic relationship between different levels of experience in a technical team. The junior developer provides the speed of agent interaction, while the senior developer provides the context of what is actually worth shipping. This balance ensures that the speed of AI does not lead to a decrease in the reliability of the final product.

Managing the flood of code

The open-source coding agent Goose serves as a practical example of how these tools affect project maintenance. Because agents make it easier for people to propose fixes or new features, maintainers are seeing a higher volume of contributions. To manage this, the project moved to an issue-first process where humans agree on a design before any code is generated.

Jones explains that the goal is to talk about the problem and the direction with maintainers before collaborating on the design. This prevents the system from being overwhelmed by low-quality patches that an agent might generate without understanding the broader goal. It ensures that the human maintainers remain the ultimate authority on the project's direction and architecture.

Even with these human-led processes, agents are still used to handle the heavy lifting of the queue. They are currently used to triage issues, review incoming code, and run security scans. These tools help maintainers manage the sheer volume of work while the humans focus on the high-level decisions that require nuanced judgment.

What to watch

  • How the MCP Associate certification will influence the hiring of developers.
  • Whether 'issue-first' processes become the standard for all open-source software maintenance.

Try it

Look into the Model Context Protocol (MCP) to see how developers are standardizing how AI agents connect to different data sources.

Sources and links

Read the original at The New Stack

First reported by The New Stack. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news