The SaaS Apocalypse Is a Goldmine With Figma’s Matt Colyer

AI Era Expands Developer Base, Creating a SaaS Goldmine

Matt Colyer from Figma refutes the common narrative of a 'SaaS apocalypse' in the age of AI. He explains that far from being a threat, the integration of AI capabilities into products presents a massive opportunity for SaaS companies.

Colyer points to a dramatic shift in the estimated number of developers worldwide. Historically, this figure hovered around 25 to 40 million, but with AI tools, he believes it could skyrocket to a billion or more. This exponential growth signifies an explosion in software creation.

For SaaS businesses with established products, this means the current period is a 'gold mine'. The surge in software development creates significantly more demand and opportunities for platforms that support and enhance this expanded ecosystem.

The ability to quickly 'vibe code' or build initial versions of applications using AI has become a mainstream expectation, particularly since January of this year. This accessibility lowers the barrier to entry for creators, further fueling the growth in software output.

if you're in that space, like it, it, it means it's a gold mine, right?

The Hidden Pain and Ongoing Costs of Custom-Built Software

Matt Colyer recounts his experience building a personal email agent two years ago, which began as a "terrible Python script." This custom solution proved unreliable and rickety, with replies sometimes failing, creating a frustrating maintenance burden.

His journey with running this self-built agent illuminated the hidden pain associated with custom software. Matt realized that software companies provide significant value beyond just writing code, managing the substantial overhead required to ensure reliability and functionality.

This firsthand insight has significantly shifted Matt's approach to tools. He now finds himself increasingly purchasing commercial software solutions, preferring to pay for services and have someone else manage agents rather than endure the ongoing costs and maintenance of building and sustaining his own applications.

I'll be honest, I'm buying more software these days than I ever did before 'cause I'm like, you know what, that, that tool seems cool, like I'm just gonna pay somebody else to run my agent for me.

Enhancing personal agents with memory and proactive daily summaries

Initially, the challenge of managing numerous emails led to the creation of a simple agent. The first version directly fed email inbox content to an LLM, often condensing multi-page emails into just a few bullet points. This basic summarization, while functional, highlighted the agent's limited capacity.

The agent then evolved to include a memory system. This addition was crucial for improving its utility beyond simple, one-off summarization. The next significant step was making the agent proactive rather than merely reactive.

A key breakthrough involved having the agent compile a summary of all relevant information and send it daily at a specific time. This proactive delivery meant the information would "show up" in the user's inbox, eliminating the need to actively seek it out in a tool. This approach, similar to concepts seen in OpenClaw, transformed the agent into a more engaged and valuable assistant.

This shift from a reactive summarization tool to a proactive information delivery system fundamentally changed how the agent served its user, ensuring critical insights were consistently presented without prompting.

the proactive part is I think the thing that really like set it on fire

AI's Challenge with Information Filtering and Overcorrection

Using AI for summarization in complex work environments presents a significant challenge: teaching agents to effectively filter information. The goal is to imbue AI with the skill to discern what truly matters from an immense amount of data, such as company meeting transcripts. It's difficult to define what constitutes relevant information, leading to AI systems struggling with context and priorities.

A common issue arises when AI models, like CodeX, are tasked with reviewing all recorded company meetings to identify relevant points for a user. While useful for gaining insights from meetings one didn't attend, the AI often provides information that isn't quite right. When corrected, the AI tends to overcorrect, delivering excessive and overly literal information.

This overcorrection means the AI gives too much of what it was told was wanted, but without the necessary nuance or filtering, turning a helpful tool into an overwhelming one. This problem highlights a key area where AI still needs development: sophisticated contextual understanding and the ability to refine its output based on subtle feedback, rather than making broad, literal adjustments.

This difficulty in teaching AI to prioritize and contextualize information, especially without overcorrecting, is considered an unsolved problem in current AI application. It impacts productivity tools designed to manage information overload, such as email or meeting summarization systems.

If it gives me stuff that isn't quite right, and I tell it it's not quite right, it overcorrects and it gives me all the things that I said I wanted, but like way too much and way too literally. I feel like this is like one of the unsolved problems at this point.

Achieving Inbox Zero with AI and Voice Commands

Dan Shipper has consistently maintained Inbox Zero for four straight weeks by leveraging AI tools. He uses Codex to draft email replies and Monologue for voice-activated interactions, streamlining his email management process.

A significant benefit of this approach is the speed of voice commands compared to typing. Speaking is roughly two to three times faster, and it also helps alleviate physical strain like carpal tunnel, offering a more efficient and comfortable way to interact with computers.

While talking to a computer can feel awkward, Dan employs a trick to make it more natural: he pretends he is screen-sharing to someone using a tool like Loom. This mental framing makes the voice interaction less strange and allows him to talk through problems effectively.

The 'audio unlock' represents a substantial yet often overlooked productivity boost. Despite the initial social barrier or perceived oddness of talking to a computer, the efficiency gains make it a powerful method for managing tasks.

it's like roughly twice or three times as fast to talk than typing.

Redefining Design with Figma's On-Canvas AI Agents

The discussion moves into how AI changes product strategy in the design world, specifically contrasting text-box interfaces with direct manipulation on an infinite canvas. While generative AI can produce a good first pass from text prompts, internal designers at Figma emphasize the necessity of physically moving elements around to refine details.

Many current generative UI experiences still default to chat-based interactions, often referred to as 'text box rules'. However, Figma is developing a new phase in its AI strategy by launching 'Agents', moving beyond this common approach.

Figma's 'Agents' allow users to interact with AI directly on the infinite canvas. This feature aims to provide a more ergonomic and intuitive design experience by enabling direct manipulation of elements, contrasting with the limitations of relying solely on text commands for precise adjustments.

typing is good for a first pass, but like to actually get the details right, I need to be able to like move stuff around.

AI Can Enhance Divergent Thinking in Design Beyond Linear Text Prompts

A core principle in design, often called the 'diamond' principle, involves two phases: divergent thinking and convergent thinking. Divergent thinking is about generating a wide array of ideas without judgment, much like a brainstorming session where every idea is valued. Convergent thinking then refines these ideas into viable solutions.

Current AI tools, particularly those based on linear text-box prompts, often limit the divergent phase. These prompts tend to guide users through a 'this, then that' process, which restricts the spontaneous exploration of truly varied approaches. This linear interaction means the human user is still primarily driving the input, guiding the AI step-by-step rather than allowing it to proactively suggest diverse directions.

Moving beyond text boxes, such as integrating AI agents directly onto a design canvas, could unlock greater divergent thinking. These agents could be tasked with taking an initial design frame and actively pushing it in many different directions, suggesting variations in grayscale, layout, or entirely new concepts. This mimics the benefit of having teammates with diverse starting points, where new creative solutions emerge from different perspectives.

Such canvas-based AI agents would not just double down on existing ideas but would proactively explore a wide spectrum of possibilities. This allows designers to generate a much broader set of concepts for a marketing page or other projects before moving to the convergent phase of refining the best options.

if we get outside of the text boxes, 'cause I think text boxes are super limiting and it's very much like a linear, like, well, this and then that and then this.

Figma's Dual Strategy for AI Agents and Third-Party Integration via MCP Server

The future of design with AI agents encompasses both divergent and convergent applications. Divergent agents help explore numerous possibilities, while convergent agents refine and narrow down options, such as generating a new landing page within an existing design system. Many designers appreciate AI for automating repetitive tasks like creating 'nth landing pages' or graphics, which are more convergent in nature.

Figma adopts a dual strategy for AI agents, supporting both internally developed agents and third-party solutions. The company believes that design and engineering workflows are increasingly blending, anticipating a future where everyone acts as a builder, approaching problems from various perspectives.

A key component in supporting third-party agents is Figma's MCP (Multi-Cloud Platform) server. This server provides a standardized interface across various tools, facilitating seamless integration. This 'code to design' approach enables common scenarios, such as updating a sign-up page to comply with GDPR requirements.

I mean, I think we embrace both, right? Like, and I think this is like, I think design workflows are different than engineering workflows, but like the lines are blurring, and so like I think in the future we're gonna be all builders, right?

Figma's MCP Server Connects Code and Design Through Agent-Powered Workflows

Figma's MCP server streamlines the connection between code and design through two primary agent-driven workflows. These workflows address common development needs, from small code-driven updates to generating code directly from designs, significantly reducing manual effort.

The first workflow, 'code to design,' helps developers translate code changes into design assets. For instance, when adding a GDPR checkbox, an agent can take the code modification and present it within Figma. This allows designers to use familiar direct manipulation tools to precisely adjust the layout and styling, removing the drudgery of manual translation.

The second workflow, 'design to code,' automates the creation of code from Figma designs. An agent analyzes a Figma design, extracting properties, components, and design library guidelines. It then uses this information to inspect the current codebase, create a new branch, generate the necessary code changes, and open a pull request, complete with a screenshot of the updated design.

For effective internal agent experiences, context and personalization are essential. Agents perform better when they have specific knowledge about the product and the user's workflow, allowing them to provide more accurate and relevant automated assistance.

It's like the agent will be like, 'Okay, cool, let me like look at your current codebase, I'll make a branch, create a PR, make the changes, and then like, you can ask the agent, be like, Okay, take me a screenshot and put on the PR.'

Personalization and Context Are Key for Effective Design Agents

Effective design agents move beyond generic functionality through deep personalization, which transforms an "okay" agent into one users genuinely love. For platforms like Figma, this personalization often comes from understanding and leveraging existing design systems, ensuring that any designs an agent creates are usable and consistent with established structures.

Matt Colyer confirms that Figma is actively exploring proactive agents, acknowledging the challenges in getting them right. The company aims to develop solutions that allow for rapid innovation driven by AI agents while ensuring the output remains consistent with an organization's core values, addressing the challenge of human eyes becoming a bottleneck in reviewing an exploding amount of software.

Colyer notes a significant internal shift at Figma, even within a short period since January. Across engineering, product, and design teams, there's been a noticeable increase in experimentation with new AI-driven ways of working, highlighting a rapid internal adoption of these emerging technologies.

personalization is often like kind of the last thing, like you get it just working for everybody first. But I think the thing that really differentiates like an okay agent to one that people really love is the personalization aspect.

Figma's AI Initiatives Transform Internal Product Operations

Matt Colyer explains how AI has fundamentally shifted problem-solving within Figma's product organization. The core insight is that every problem, when viewed through an AI lens, becomes a 'context problem.' The actual work then revolves around effectively framing the problem with the correct and comprehensive set of information.

Figma's product operations team capitalized on this by aggregating structured data from various internal tools like Asana, Slack, and GitHub. They recognized that much of their work involved structured data, and connecting these sources laid the groundwork for AI-driven solutions.

A key example of this transformation is an AI 'skill' designed for onboarding new team members. When a new person joins, the AI agent, provided with just the new hire's identity and team, can access the org chart, identify the team's 'trifecta' (product, engineering, design leads), and even research and suggest relevant Slack channels. This process, driven by the right contextual data, creates a customized onboarding file automatically.

This AI-powered system provides an uncannily good starting point for new hires, demonstrating how providing structured context to AI agents can significantly streamline traditionally manual and time-consuming operational tasks within a company.

you kind of realize every problem becomes a context problem, and it's all, like, then the work becomes about like framing the problem with the right set of like information.

Local AI Agents Excel by Accessing Full Personal Context

Local AI agents like Claude Code and Codex are proving highly effective because they operate directly on a user's computer, granting them comprehensive access to personal context. Unlike cloud-based agents that require manual connections to various data sources, local agents can leverage all the information a user has access to, significantly enhancing their capabilities.

For example, Dan asked Codex to suggest recipients for a newly published article. The agent seamlessly went through his emails and texts, identifying five relevant individuals he might have overlooked. This ability to scan and utilize personal communication history demonstrates the power of having direct access to a user's complete digital environment.

The underlying AI technology has long possessed the capability for such tasks, provided it had sufficient context. The recent breakthrough isn't in the AI itself, but in the development of the right 'harness and form factor' that allows these agents to operate locally, access vast amounts of personal data, and perform tasks more independently than previously possible.

This approach addresses the challenge of connecting AI to fragmented personal data, a problem that even large tech companies like Apple are exploring with concepts like Apple Intelligence, which aims to leverage the rich personal data available on mobile devices.

The technology, like AI itself, if you gave it all the context, would have been able to do this for a while, but it's only now that it, it can, it, it's in the right harness and form factor and it can do it a little bit more independently than it was able to before.

Apple and Google's Contextual Advantage in the AI Agent Race

Even if they appear behind, Apple and Google are uniquely positioned in the AI agent race due to their status as "kings of context." Both companies possess vast amounts of user data, providing them with a foundational advantage for developing truly intelligent, always-on agents. This deep access to personal information allows for a level of personalization and understanding that other AI developers may struggle to replicate.

Google is actively leaning into this advantage, as seen at Google I/O, by integrating its AI products more deeply with user content. There is anticipation for a product like Spark, an always-on agent designed to connect with and manage all Google content, potentially automating tasks like email inbox management. This move signals a strategic shift towards leveraging their existing data ecosystems.

Apple also benefits significantly from its strong hardware foundation. Historically, Apple's superior hardware has compensated for occasional lags in software, giving them ample time to refine their AI offerings. Their strategy is also smart from a privacy perspective, addressing user concerns about uploading personal data to the cloud, and reinforcing their position in the market.

I think it still matters because I think even them being late to the game, they are still the king of context, right?

Human Review Becomes the Bottleneck for AI-Generated Content

As AI capabilities increase, the challenge in product development shifts from content generation to content review. AI agents are now readily available and cheap enough to produce a vast amount of material, but human capacity to review and approve this output has become the limiting factor.

Users are being inundated with new AI-generated content, leading to an overload where people struggle to process and decide what to do with the volume of information. This isn't just about summarizing existing content; it's about new creations requiring human oversight.

Solving the problem of scaling review processes is critical for fully leveraging AI's potential. The industry is actively exploring various approaches, such as new presentation formats like video walkthroughs or screenshots, or even employing other AI agents for automated review, though the optimal solution is still being developed.

The challenge is not limited to a single company; many organizations are simultaneously trying to understand how to design workflows that effectively manage the influx of AI-generated content and integrate it into existing product development cycles.

Foundational Skills and Curiosity Remain Essential for Product Managers in the AI Age

Matt Colyer addresses concerns about career progression for product managers and designers in an AI-driven industry, arguing that foundational skills are still crucial. He uses the analogy of learning long division and calculus in school despite having calculators; while not used daily, these fundamentals provide a necessary understanding of how things work.

Colyer differentiates between two approaches to using AI tools. One involves simply accepting outputs, such as asking ChatGPT for a bubble sort algorithm. The more valuable approach comes from a curious person who would take that bubble sort, convert it to assembly, and then ask for an explanation of registers and L1/L2 cache, dissecting how the solution is truly built.

He emphasizes that curiosity is the most vital trait for professionals. Individuals who merely accept AI outputs will struggle to leverage new tools effectively. In contrast, those who actively push boundaries, understand the underlying mechanisms of AI, and possess an inherent curiosity will be the ones who invent the next generation of tools and maximize the potential of existing ones.

I think the most important thing is to be a curious person, because I think with these new tools, the people who can't leverage them are the ones that just kind of accept the output, and the people that are able to invent the next set of tools and really drive the tools to their maximum are the ones that are pushing the boundaries and understand how it's put together.

How Large Language Models Fuel Curiosity

Cultivating a curious mindset is crucial for truly understanding how things are put together and how they actually work. This deep dive into mechanisms, rather than just seeking quick answers, is key to leveraging new technologies effectively and enhancing personal understanding.

Large language models (LLMs) serve as a powerful tool for this inquisitive approach, almost like a real-world manifestation of the 'Hitchhiker's Guide to the Galaxy'. They provide immediate access to information, making it easier to explore any topic that piques one's interest.

For instance, even simple questions like 'What is a squirrel?' can be answered instantly by an LLM, allowing users to quickly gain foundational knowledge and then delve into deeper, related topics. This interactive exploration makes learning an engaging and enjoyable process.

This ability to quickly satisfy curiosity about diverse subjects turns learning into a more fun and catnip-like experience, encouraging users to continuously seek new insights and build a more comprehensive understanding of the world around them.

I feel like LLMs are the book. Like, it is literally the manifestation of it.