Figma's Strategy for High-Quality AI Agent Adoption

My name is Al Bloom. I am a software engineer at Figma and in my talk today we're going to talk about how we've adopted or are adopting soft agent into our workflow at Figma while maintaining high quality for our code base. Eyal Blum, a software engineer at Figma, discusses the company's shift from traditional software development to incorporating AI-first tools. His presentation focuses on the internal engineering organization's workflow and how AI agents are being integrated without compromising code quality.

There's kind of a three-act process of AI adoption. You start with picking up something... Then you start applying those same practices to bigger problems and AI fails pretty badly at that... And then from that point you start building the real skill which is learning how to use AI correctly and put the right guardrails. The AI adoption process within an organization unfolds in three acts: initial success with simple tasks, followed by a phase of significant failure when these practices are applied to more complex problems, leading to bugs and a loss of trust. The third act involves developing genuine skill by understanding how to use AI correctly and implementing appropriate guardrails and contextual awareness.

The adoption is uneven. We have teams that are very AI forward and have already transformed their entire workflows and then we have teams that are still experimenting in the earlier act and or have lost confidence and they all need to work together in order to ship our product. Figma observes an uneven pace of AI adoption among its teams; some are highly advanced, having fully integrated AI into their workflows, while others are still in experimental stages or have lost confidence due to previous setbacks. Despite these varying adoption levels, all teams must collaborate effectively to ensure the continuous shipment of their product.

One thing that we've heard a lot from developers and managers have been noticing is a reduced developer agency causes engineers to lose some of their job satisfaction. If a lot of people used to take a lot of pride and enjoyment in writing code and getting into the flow, and a lot of people feel like that's been lost... getting into more of a prompt cycle where they just wait on output from AI and then speak to the AI that's like not as much fun as they used to have and they be getting burned out. The transition to AI-assisted coding has led to reduced developer agency, impacting job satisfaction for engineers who previously found pride and enjoyment in writing code and achieving a 'flow' state. This shift towards a 'prompt cycle,' where developers largely wait for and interact with AI output, is causing burnout and a sense of lost fulfillment among the engineering staff.

Our best engineer, the one that hold all their context in their brain, um they end up getting a lot of the burden... They are like holding together with with like their mental duct tape all the places that agents are not working well and they're preventing all the really bad stuff from coming in... they get so much burden and and become bottlenecks and gets really frustrated. The most experienced engineers, those with the deepest institutional context, often bear the heaviest burden in manually verifying AI-generated code. These individuals become critical bottlenecks as they prevent errors and risks stemming from AI agents, leading to significant frustration and making them the slowest to adopt AI due to their acute awareness of potential issues.

All of a sudden all the design docs and all the Slack messages and all the emails have gotten three or four times as long and we've gotten two or three times as many emails. And they say basically as much as they did before. So, communication has gotten quite inefficient. The introduction of AI has led to a significant increase in the length and volume of communication, including design documents, Slack messages, and emails, often becoming three to four times longer and two to three times more frequent. This surge in communication, however, has not correlated with a proportional increase in meaningful information or shipping speed, resulting in overall communication inefficiency.

Investing in verification is probably the highest value thing we can do in our code base... when Playwright MCP came out, instead of having humans navigate the code, now the agent can explore the code. That was a big win unlock for productivity... It even better if when you find something that the agent has found to be useful, take the time to take that and encode it into a deterministic flow. Eyal Blum emphasizes that investing in verification, particularly through a 'Left-shift' approach, is the most valuable activity for their codebase. He cites examples like Playwright MCP, which allows AI agents to explore code autonomously, greatly enhancing productivity. Furthermore, he advocates for encoding successful AI patterns into deterministic flows to reduce token usage and improve overall reliability.

if you tell your skills or your agent to write the code that you're writing, um like at the red to ring to red to green at the TDD style, it almost always gives you better results. Applying Test-Driven Development (TDD) principles to AI agent-driven coding consistently yields superior results. This approach, which involves defining tests before writing the code, encourages agents to meet specific verification criteria rather than adapting flawed code to pass existing tests.

move as much as you can down to the deterministic analysis, whether it's linting, the compiler, and the unit test themselves. Eyal Blum recommends shifting as much testing as possible to deterministic analysis, including linting, the compiler, and unit tests. Human review, therefore, should be reserved primarily for architectural functionality and high-level strategic goals.

Spending a lot of time writing the plan and then selling enough to the agent basically as an implementation that can be done automatically is something that we find to really kind of reintroduce the joy of of building back into the process. Eyal Blum argues that detailed planning, rather than pure prompting, can restore the intrinsic joy of coding for developers. He suggests that dedicating a week to meticulous planning can lead to massive efficiency gains, as these plans can be reviewed iteratively by team members before AI implementation, ensuring quality and alignment.

very important to start with the why at the top. It really helps preventing agent drift. Eyal Blum stresses the importance of starting a plan with a clear 'Why' in an executive summary to prevent AI agent drift. He also recommends breaking down plans into small, independently verifiable segments, using 'coffee-break logic' to determine the appropriate size and scope of each task.

having kind of a validation gate or an exception criteria for each phase really helps um make the plan resilient to drift. Implementing a validation gate or exception criteria at each phase of a multi-stage plan is crucial for making it resilient to drift, according to Eyal Blum. This approach prevents errors from propagating downstream by catching and addressing unvalidated assumptions early in the process.

there's diminishing return in trying to centralize everybody on one thing, but as long as it works for their flow and other people can iterate with them, I find that it generally works very well. Eyal Blum asserts that attempting to centralize all teams on a single AI workflow yields diminishing returns. Instead, he advocates for empowering developers to configure tools that best suit their individual workflows, provided these setups allow for effective collaboration and iteration with other team members.

This is in the pre-AI world, this plan I probably worked on it for a week. I aligned with other with three other teams for another week on that, and then I just sent it to an agent to implement overnight, and it came back. Eyal Blum illustrates a significant efficiency gain, reducing a task that would have taken six weeks of manual coding to just one week. The process involved a week of detailed planning, followed by another week of alignment with three other teams. After this, the plan was submitted to an AI agent for overnight implementation, demonstrating a substantial acceleration in development.