"Model capabilities are already sufficient—if we're going to compete, compete on infra" | Interview with Guanlan Dai: Founder of Runta
👦🏻 Podcast Interview: Koji
🥷 Edited by: Crossing
🧑🎨 Layout: Zeoooo
🚗This week on Crossing, our guest is Guanlan Dai, Founder/CEO of Runta. Runta is a Silicon Valley star startup in the Agent Infra space, having just raised a $20 million Seed round led by a16z; Jeff Dean and Fei-Fei Li also participated as individual angel investors.
Before founding Runta, Guanlan spent over a decade in cloud computing and infrastructure: he was a core technical lead for Cloudflare Edge, then joined Kong as an early core member and Engineering Director, leading the development of cloud-native and AI gateways.
His reasoning for starting a company now is straightforward: model capabilities are already good enough; if there's going to be competition, compete on infra — there will inevitably be more agents than humans, and the key questions in AI will become: where do they run? How are they managed? Who's responsible when something goes wrong?
In this conversation, we start with the sharp turn in token culture: in just six months, token maxxing (the more you use, the more glory) rapidly shifted to token minimizing (ROI, budgets, attribution, cost-cutting). What signals are worth paying attention to behind this shift?
You'll also hear about:
- When software execution shifts from "logically clear determinism" to "LLM probability": how is AI shaking the very foundation of the digital world?
- Who will build the "utilities" of the AI era? And why, this time around, do startups actually have a better shot than giants?
- What does a world with billions of agents working simultaneously look like? Can today's internet and cloud even handle it?
- Why do humans, click after click on "Allow," unknowingly surrender boundaries and control?
- Why is it said that "in the next 10 months, there will inevitably be a catastrophic incident"? And how will that incident "re-educate" everyone?
Listen on WeChat:
Listen on Xiaoyuzhou:
🎬 The video podcast is also available on Koji's WeChat Channels, Xiaohongshu, Bilibili, YouTube, and other platforms.
🎪 Crossing will host Guanlan Dai for an online closed-door sharing session on Agent Infra on August 17 (Sunday), diving deeper into topics not fully covered in the podcast. Engineers and entrepreneurs interested in the Agent Infra space are welcome. To ensure quality discussion, attendance is limited to 20 people. Scan the QR code below to register.
Rapid-Fire Q&A
👦🏻 Koji
Let's start with the traditional rapid-fire round to help everyone get to know you quickly. Guanlan, how old are you?
👨🏻💻 Guanlan
👦🏻 Koji
Where did you go to school?
👨🏻💻 Guanlan
Northeastern University.
👦🏻 Koji
Your MBTI and zodiac sign?
👨🏻💻 Guanlan
Libra. My MBTI oscillates between INTJ and ENTJ, though I personally lean more toward I.
👦🏻 Koji
One sentence to describe your current company and product.
👨🏻💻 Guanlan
We believe there will be more AI agents than humans in the future. Faced with these billions of agents — where do they run, and how are they managed once running? Runta is the execution substrate built for these AI agents.
👦🏻 Koji
Current team size?
👨🏻💻 Guanlan
We're about 10 people on the team, backed by hundreds or thousands of agents doing the work.
👦🏻 Koji
One sentence on what you did before starting up.
👨🏻💻 Guanlan
I previously worked on Cloudflare's Edge cloud platform, and on gateways and cloud at Kong. Altogether roughly ten-plus years, always building infra for the software and SaaS era.
Runta Raised $20M in Seed Funding — How Did You Get Jeff Dean and Fei-Fei Li as Individual Investors?
👦🏻 Koji
Runta's Seed round raised $20 million, led by a16z, with Jeff Dean and Fei-Fei Li participating as individuals. I'm curious — how did you get them on as angel investors?
👨🏻💻 Guanlan
First, everyone probably knows Jeff Dean — he's a legend in distributed systems and systems engineering, having lived through the infra evolution across Google, DeepMind, and various eras.
I was fortunate to have a very deep conversation with him. He asked me a question that left a deep impression: "Because LLMs have made the most fundamental execution unit of computer systems probabilistic, how should the infra layer be built?"
This is actually a very hard question to answer. Because over the past few decades, systems engineering concepts like retries, idempotency, transactions, and recovery have all been built on the assumption that software is deterministic. The same input should produce the same output. If there's variation, it must be an enumerable exception. His question was essentially asking: if the foundational stone at the very bottom disappears, how does the building above stand?
At the time, my thinking was that uncertainty is indeed the fundamental nature of large models themselves. But once agents truly enter production environments, the complexity of distributed systems actually increases. Needs around recovery, isolation, forking, and so on all rise dramatically.
If the underlying infrastructure can't handle these things well, it forces everyone to push all that complexity into the runtime layer to solve.
In the previous generation of infra, we solved how physical machines run, how resources are isolated, so developers don't have to care about hardware. In this generation, I believe the most core and difficult problem has become how execution itself works.
I've exchanged many ideas with Jeff Dean and Fei-Fei about how this generation should iterate. At the root of it, the execution layer has now become the hardest part of the entire AI systems engineering stack.
Agent run cycles are getting longer and longer. Before, it might have been a few minutes or a few turns of chat; now it's routinely hours or even days for an agent to run. In these scenarios, traditional infra simply can't handle isolation, forking, live migration, and other high-difficulty problems well. The challenges here are enormous.
Also, I think a main reason they invested is that I previously participated in building the previous generation of infra from first principles at Cloudflare and Kong.
👦🏻 Koji
How did that experience connect to today?
👨🏻💻 Guanlan
When we started, Kubernetes and containers didn't even exist yet. We had to derive from first principles and the most fundamental, bottom-layer needs what a system should look like, and witnessed it eventually become a cornerstone of the modern internet.
At Kong, we processed hundreds of billions of API requests daily — so at that scale, even a 0.00001% error probability meant inevitable failures in daily operations.
I've lived through and handled that magnitude of systems complexity. This ground-level practical experience gave them confidence in us.
At the time, Jeff Dean actually wanted to invest more, but I said there simply wasn't enough allocation left. Turning down Jeff Dean's additional investment is, in retrospect, quite an interesting experience.
Why Did a16z Rarely Write a Dedicated Blog Post for a Seed-Stage Company?
👦🏻 Koji
a16z invests in many companies, but it's rare to see them invest in such an early-stage team and then deliberately write a blog post about it. They specifically wrote "Why We're Investing in Runta." Why do you think they placed such emphasis on you?
👨🏻💻 Guanlan
Indeed, they rarely do this for seed rounds. The article was personally written by partner Martin, because he saw this not as a feature or product, but as a platform-level generational opportunity.
In his view, this is a generational shift in compute. Just as we went through waves of evolution from client-server architecture, to data centers, to virtualization, to containerization — each wave produced enormous generational change.
In every previous generation, what we hosted and managed was static software and services. This time, the underlying hosted objects have become agents. If this logic holds, it's an enormously large platform-level opportunity. That article breaks down this logic very thoroughly — highly recommended reading.
👦🏻 Koji
In one very simple sentence, explain what are the most critical problems Runta solves for customers?
👨🏻💻 Guanlan
We work at the deepest layer of infra. We're adding deterministic control and management into a non-deterministic, probabilistic workflow, helping enterprises actually put agents to work.
For example: where does the agent run? How do you guarantee its security permissions while working? When it needs to access customer data, touch production environment permissions — at the moment it touches, we need strict isolation, control, and audit through the execution layer, and support it to scale.
That's what Runta does.
Why Did Token Maxxing Shift to Token Minimizing Within Three to Six Months?
👦🏻 Koji
Let's start with a major shift that's been happening across AI and the Agent space. A few months ago, everyone was talking about Token Maxxing — companies encouraged employees to burn through as many tokens as possible, even wearing it as a badge of honor. But lately the narrative has flipped completely, and now it's all about Token Minimizing.
What's the backstory of this reversal, from what you've observed?
👨🏻💻 Guanlan
At first, when reasoning models like o1 came out, people discovered that if you gave large models a chain of thought or enough thinking steps, they became incredibly smart — worth the extra token cost even.
Gradually, this logic got extrapolated into an organizational management tactic: the more tokens an employee consumed, the more "AI Native" they were. It became a measurable KPI. Hence Token Maxxing, with companies aggressively pushing the transformation — especially giants like Meta and Amazon.
Late last year, when I visited some cutting-edge customers and AI labs, I even saw monitors mounted on office walls cycling through who had consumed the most tokens that week, with rewards for top-ranked employees.
At the extreme, companies like Meta and Stripe directly tied token usage to year-end performance reviews. That's when things really went off the rails.
👦🏻 Koji
What's the most extreme example of how this incentive got distorted?
👨🏻💻 Guanlan
The most extreme cases I heard about were employees using the most expensive SOTA models for even the simplest daily tasks. Some even deliberately wrote meaningless infinite loops in their code just to rack up token metrics.
In March this year, Jensen Huang said in a speech: "An engineer making $500,000 a year who doesn't burn $250,000 worth of tokens isn't good enough." Looking back now, that wave definitely got overheated.
From aggressively encouraging usage to companies slamming the brakes on budgets — the reversal took just three to six months.
👦🏻 Koji
What's the fundamental reason for this dramatic reversal?
👨🏻💻 Guanlan
People started recalculating ROI. Especially after Uber's CFO publicly stated: "We burned through our entire annual LLM token budget by April — but what did we actually get for it?"
First, the ignition phase of awareness is over; AI literacy education is largely complete. Second, tokens are genuinely expensive, and enterprises had massively underestimated the cost of deploying models at scale. Combined, these factors made everyone realize Token Maxxing wasn't sustainable.
👦🏻 Koji
You're also a founding team yourselves, about ten people. How do you internally manage rules around employee token and API usage?
👨🏻💻 Guanlan
We started by hiring a lot of traditional infra people and just gave them unlimited tokens — creating this illusion that tokens were as free as water.
Now we've introduced a tiered system. I subscribe the team to several top-tier Ultra Plans for everyday unlimited use. If those quotas run out, they need to submit a request to me explaining where those tokens went.
It's really about introducing a tiny bit of friction into the process, making people aware this isn't free, so they'll spontaneously think about purpose when making calls. But overall, I still keep very generous room for the team to experiment and fail.
Why Give Up Your Position at Kong to Become a CEO for the First Time?
👦🏻 Koji
You worked at Cloudflare and Kong for a long time. What stages were these companies in, and what specifically did you handle?
👨🏻💻 Guanlan
I was among the earliest engineers to join, and served as the technical lead for Cloudflare's core Edge cloud. I was fortunate to participate in building edge caching, WAF (web application firewall), and the early Worker system (edge computing platform).
The core problem I solved was how to build edge networks for software services, helping clients access the nearest node faster while providing gateway-level security and acceleration.
At Kong, it was similar — enterprise gateways. When you order takeout or call API services, how do you do rate limiting, routing, and traffic direction? At Kong, I worked extensively with Fortune 500 clients, helping them build microservices platforms and execute digital transformation.
👦🏻 Koji
So this Runta venture marks your first time as CEO. What drove that decision?
👨🏻💻 Guanlan
Serving enterprise customers and developers over the years, I found everyone desperately craving Agent deployment. Large models convert electricity into tokens; an Agent's job is to convert tokens into real business value.
But Agents have an essential property: unlike traditional software, they execute probabilistically. Every run is different — even the code they generate varies. You simply can't exhaustively cover them with fixed rules.
At the same time, they're different from humans too. When humans mess up, someone takes the fall: if Koji makes a bad investment, LPs come for you, your personal reputation takes a hit; if a system I managed went down, the CEO came for me. Human society has a "accountability force" that constrains responsibility. But when an Agent fails, who's responsible? The LLM vendor? The Agent developer? The enterprise? Nobody can say.
It's like software, and like a person, yet not fully either. I started Runta because I believe traditional runtime architectures are fundamentally unsuited for this new workflow. There's a massive infrastructure gap here.
Cloud and SaaS-era infra was designed for deterministic software. Facing completely non-deterministic Agents, what should the execution substrate look like? That question fascinates me, and it's the core reason I gave up my executive package at Kong to start this company.
👦🏻 Koji
If people are using Codex or Claude Code, can they use Runta? Or do they need to open-source deploy their own Agent for it to be useful?
👨🏻💻 Guanlan
They can use it — we support any coding Agent. We're a fully open execution platform.
For example, when you use Claude Code to check your email. Which emails should it read? What shouldn't it access? How do you manage its databases, key credentials? Put it on our execution layer, and you get natural elasticity and comprehensive governance.
We provide a highly customizable, cloud-hosted virtualization platform. We have our own custom system kernel and network mechanisms. We place zero restrictions on what specific Agent you run.
👦🏻 Koji
From your current customer conversations, what do they find most valuable about Runta?
👨🏻💻 Guanlan
Particularly this past month, budget and Token Maximizing and compute spending are top of mind — where Agents run, how to run them more elastically and scalably when they run frequently.
Then there are financial sector clients with strong governance requirements. They really want Agents doing real work to differentiate from competitors, but they're equally worried about uncontrolled scenarios once Agents start running.
A recent security study found that among Fortune 500 companies using Agents, most have already leaked confidential or customer information. That's a massive problem.
Why Is Agent Infra More Likely a Startup Opportunity Than a Public Cloud Business?
👦🏻 Koji
Why is this a startup opportunity? Can't the public cloud giants do this?
👨🏻💻 Guanlan
Existing public clouds were entirely designed for the previous generation of SaaS and traditional software. From foundational architecture to resource allocation to billing models, the baggage is heavy.
It's very similar to the GPU explosion. The previous generation of data centers was built for CPU. When this GPU-centric wave hit, traditional data centers had to face fundamental reconstruction, giving rise to numerous Neo Clouds.
Public clouds will certainly follow, but their own interests and existing technical systems are in conflict. Opportunities for disruptive innovation tend to favor startups that can travel light, without historical baggage.
What Differentiates Runta from Anthropic and OpenAI?
👦🏻 Koji
Beyond public clouds, major model vendors like Anthropic and OpenAI are also building their own Agent Infra. What's your differentiation against them? Anthropic has already released Managed Agent system services, for example.
👨🏻💻 Guanlan
Indeed, every investor asks us this.
But from the logic of actual enterprise deployment, no company will deeply bind its business to a single model or single cloud service. Multi-model, multi-cloud, risk distribution, and cost control are enterprise necessities.
Moreover, for large model vendors, their core mission will always be improving base model capabilities. Building applications and infra is primarily to gather enough signals to improve the model.
So base models and infra execution substrates are fundamentally different in industry division of labor.
👦🏻 Koji
Then where do you think your biggest potential competitors will come from?
👨🏻💻 Guanlan
I actually think it's the actively transforming Neo Clouds or previous-generation serverless giants. They've been through the cloud migration wave, the foundation is still there.
But again, for them this is just one of many business lines. For us, this is our sole 24-hour focus. A focused team can see deeper into the fundamentals.
Why Can't E2B and Daytona-Style Sandboxes Support Long-Horizon Agents?
👦🏻 Koji
There are some teams focused on Agent Infra like E2B and Daytona, mainly doing Sandboxes. What's your competitive dynamic and differentiation with them?
👨🏻💻 Guanlan
Sandbox technologies like Firecracker actually originated from the previous generation of serverless (think AWS Fargate), designed for short-lived tasks — jobs that wrap up within minutes or an hour, providing brief code isolation and returning a result. For quickly running a snippet of code, this approach works great.
But agents are evolving from quick-hit tasks to increasingly long-running processes. Traditional serverless sandboxes hit severe technical bottlenecks when facing agents that run continuously for days or weeks — dynamic migration, elastic GPU allocation, and dynamic memory contraction all become major challenges.
Solving these problems requires more than patching old technology. It demands rebuilding an entire system foundation purpose-built for agents, starting from first principles and working upward.
These are two fundamentally different underlying logics.
👦🏻 Koji
The company hasn't been around long. What's the scale of your largest order so far?
👨🏻💻 Guanlan
We already have orders in the mid-six-figure USD range, but at this stage revenue is absolutely not our core metric. We're focused entirely on building a general-purpose platform from the bottom up.
We discussed the build-to-order versus standard product question before — I'd love to hear your take, Koji.
👦🏻 Koji
My instinct is to build standard products, because only standard products achieve true scale financially and operationally. Non-standard usually means every customer requires a massive custom service organization, which becomes painful to manage.
Though if a single non-standard customer brings a $50 million contract, four or five customers gets you to $200-300 million ARR — that's absolutely worth building dedicated teams for. So I'm curious how you see this playing out?
👨🏻💻 Guanlan
My view is that right now we should go after the most agent-native customers, because they represent the cutting-edge methodologies and productivity.
At the same time, we maintain periodic sync-ups with Fortune 500 companies to understand where traditional enterprises are in their deployment, what specific pain points they have, and what their red lines are.
So we're "grabbing with both hands," but our current priority and the most tech-receptive audience is that native cohort.
👦🏻 Koji
Which companies do you mean by "agent-native"?
👨🏻💻 Guanlan
Mainly two categories:
First, agent builders — developers who build agents for others to use;
Second, vertical agents — specialized companies focused on video editing, marketing, and similar verticals. We choose them because they inherently have massive business volume.
Enterprises claim to have hundreds of agents — where exactly do you draw the boundary?
👦🏻 Koji
How do you define an agent's boundaries? If I just assign a new task in Codex, would you call that an agent?
👨🏻💻 Guanlan
There's genuinely no industry-standard answer right now. Current definition logic mainly falls into two camps: anthropomorphization — giving it an identity and assigned role; and functionalization.
Behind this lies a whole set of technical challenges. Take agent permission management — we can't grant unlimited permissions, we need dynamic allocation where permissions are immediately revoked after task completion, with extremely granular management approaches.
How does AI encroach on human permission boundaries?
👦🏻 Koji
Sometimes I need my Codex to read my Gmail. Initially I felt uneasy and would set permissions to auto-revoke after an hour. Now I've just opened everything up for read-write access, choosing to unconditionally trust it won't mess around.
👨🏻💻 Guanlan
That's classic audit fatigue.
Permission management actually requires hard rules — like "read-only, no sending," or certain confidential folders being absolutely off-limits. If you let an agent send emails directly, what kinds would you have it send?
👦🏻 Koji
I use it to find birding guides. I'm into birdwatching, and when I travel somewhere new I need to connect with local professional guides.
I have Manus draft and send emails directly to reach out. I've sent several emails like this already.
👨🏻💻 Guanlan
How do you confirm every email it sends is appropriate? Do you audit it regularly?
👦🏻 Koji
At first I required it to only write drafts, which I'd review before sending. Once trust was established, I just let it send on my behalf.
👨🏻💻 Guanlan
But what if it suddenly sent an email to your LP saying you want to quit or the carry is too low — would that strike fear?
Why catastrophic agent incidents are inevitable
👦🏻 Koji
That hasn't happened yet. But I feel like once something like that occurs, even if not to me personally — a disastrous email sent out by someone else — I'd become extremely cautious and careful afterward.
If I wanted to use Runta to manage these permissions myself, is your platform open to individuals?
👨🏻💻 Guanlan
We currently have an open-source project called ClawShell. It includes local support for Lobster and Hermes. If you're running locally, you can use this open-source product to manage your email.
But on the Runta platform, what we do is far more complex — database connections, code repository read-write control, and isolated auditing.
I believe similar security incidents are inevitable in the future. When the first catastrophic case caused by agent permission loss occurs, the industry will hit a massive inflection point of reflection.
Traditional cybersecurity relies on scanning and matching — we don't do that. What we do is governance — constraining agent behavior within secure sandboxes, with complete auditability and recoverability.
Only with such a sufficiently secure and deterministic foundation can enterprises confidently grant agents higher-level permissions.
👦🏻 Koji
Currently, when you engage with enterprise clients, do you need to spend significant effort convincing them to pay for security infrastructure?
👨🏻💻 Guanlan
Actually, quite a few enterprises currently trade security for convenience.
Everyone has underlying fear, but there simply wasn't a sufficiently usable execution foundation on the market to help them balance security and convenience. As the major wave rolls in, this awareness and necessity will explode rapidly.
👦🏻 Koji
It feels like everything you're doing is preparing for that day?
👨🏻💻 Guanlan
Our vision is to help enterprises confidently hand true non-read-only production permissions over to AI agents.
👦🏻 Koji
So your entrepreneurial state is different from your previous two ventures. Before, the problems were clearly defined; this time you're taking half a step forward through ambiguity, building necessary infrastructure for what's coming?
👨🏻💻 Guanlan
Actually this market has already taken off. Everyone is now clear that purpose-built infrastructure for agent runtime characteristics is essential, especially on the execution side — it's in an explosive state right now.
👦🏻 Koji
Name one or two signals that make you feel it's taking off?
👨🏻💻 Guanlan
We currently have many customers queued up wanting our services. Because we're prioritizing delivery capacity and resources toward native teams, we're actually slowing down and temporarily turning down some orders.
👦🏻 Koji
Have you abstracted out any standardized products at this point?
👨🏻💻 Guanlan
We have some relatively standardized products, like our execution platform and sandbox scheduling.
There's also an interesting feature we've been using recently. On top of the execution plane foundation, you can run any harness — whether Codex, Pi, or something you built yourself. Once running, we can analyze whether your harness is wasting tokens, where the waste occurs, and how.
We can also help generate prompts for your harness to iterate, telling you which aspects need optimization, forming a closed loop that shows you how to run your harness "better, faster, and more economically."
Runta does 95%+ of development via vibe coding — where are human engineers still irreplaceable?
👦🏻 Koji
When you build products or serve customers now, what percentage is vibe coding?
👨🏻💻 Guanlan
Over 95%. The work of basic programmers is now completely taken over by agents.
But this actually places greater demands on human architectural capability. API design, component coupling, insight into distributed systems pain points — this work involving infrastructure experience and foundational architectural judgment can still only be done by humans.
👦🏻 Koji
Using agents to write code that governs other agents — isn't this logically risky in itself?
👨🏻💻 Guanlan
Absolutely risky. So this requires strong foresight — how to judge whether it remains valid within its effective window.
Moreover, we ourselves use our own products deeply every day, running business with agents in our own foundation. We're using our own products for self-iteration.
When agent security fails, is the root cause more likely the model, the harness, or the execution foundation?
👦🏻 Koji
In your view, if security issues emerge in the future, are they likely to be model problems, harness problems, or where would they most likely originate?
👨🏻💻 Guanlan
First, large models will certainly keep training on safety. Take the recent Fable model — once it detects sensitive information, it automatically downgrades to a weaker defensive model.
But as long as the model doesn't change its Transformer architecture, it remains fundamentally a probabilistic next-token prediction system. It simply cannot distinguish whether what you feed it is data or control instructions. Even if you insert special delimiter characters during transmission, these can be bypassed through various means.
So model security vulnerabilities cannot be fundamentally solved on the model side alone.
This is precisely why the infrastructure side must have a presence. We need to control the surrounding network, storage, and virtualized operating system, wrapping this non-deterministic "brain" with an entire "deterministic environment."
Recently, one of OpenAI's models hacked into Hugging Face to grab reward signals. Hugging Face wanted OpenAI's help investigating, but the big model labs considered it classified and sensitive, so they just walked away.
In the end, they used an open-source model to figure out what happened. Even though that open-source model wasn't the most powerful SOTA, it was controllable.
👦🏻 Koji
We're talking about open source versus closed source at just the right moment. The competition between them has been fierce lately. Besides OpenAI and Anthropic, it feels like everyone else has formed an "Avengers Alliance," sparking a new movement: stop using closed-source models, they're expensive, your data gets used for secondary training, the value proposition is poor — use open source instead.
What's your take on this, and how do you see the competition playing out over the next few years?
👨🏻💻 Guanlan
The climb of open-source models has been staggering. Moonshot AI has recently broken into the top tier, and DeepSeek's Flash model releases have all proven that open-source forces can achieve massive progress on speed, quality, and cost savings.
The United States still holds enormous leads in compute, so they've been able to attempt more frontier, high-difficulty exploration — it's a seesaw competition.
We're currently serving some companies that use Agent AI to run Auto Research for improving model training, which is quite interesting.
"Model capabilities are already sufficient" — why is the next battleground Infra?
👦🏻 Koji
If model differences keep narrowing, won't they become low-margin commodities like electricity and water? That's the biggest concern people have about the sky-high valuations of large model companies.
👨🏻💻 Guanlan
The macro trend will definitely get there, and faster than we imagine. For us developers, using Fable 5 versus Opus — in many everyday scenarios, there's no fundamental generational gap in experience anymore. The performance climb of large models is already slowing.
I'll make a bold claim here: model capabilities are actually sufficient already. If we're going to compete on anything in the next phase, it's Infra.
We have the models, but the infrastructure and harness supporting them hasn't kept up, so large-scale deployment keeps getting delayed. This is the fundamental reason we built Runta.
👦🏻 Koji
How competitive is the Infra track you're in?
👨🏻💻 Guanlan
Those frontier labs each have dozens or hundreds of top-tier teams working on similar things, but their underlying purpose is usually to improve their own models' capabilities.
On the execution substrate front, people have seen the need, but haven't fully reached consensus on "whether an Agent execution substrate is actually necessary."
I hope that a year from now, people stop debating whether an execution substrate is needed. Just as data lives in databases, Agents necessarily need to run on an Agent execution substrate.
I came on this show hoping to push industry awareness forward on this.
I hope it gets more competitive in here — only with solid infrastructure can Agent deployment truly succeed. If you're interested in this, we'd love to have you join Runta.
👦🏻 Koji
Are you mainly in San Francisco now? Or do you have offices elsewhere too?
👨🏻💻 Guanlan
We're in both Singapore and San Francisco.
👦🏻 Koji
Any career advice for engineers just starting out?
👨🏻💻 Guanlan
I think when you're just starting, you can begin by Token Maximizing.
When you're new to the workforce, having no baggage is actually your greatest asset. At the same time, maintain respect for what came before, and build up your architectural thinking and foundational reasoning skills.
If any of you are in this position, I'd be very happy to chat and explore future directions together.
👦🏻 Koji
What were the most critical choices you got right in your career?
👨🏻💻 Guanlan
First was recognizing when the wave was coming.
For example, why did I join Cloudflare? Because I saw edge computing rising, the migration from mainframes to cloud — the wave was just forming. There happened to be a startup exploring this direction with solid tech, and this wave was worth riding.
Joining Kong was because I saw the microservices and cloud-native wave. Around 2020, K8S, Mesos, and Docker Swarm were still battling it out, but everyone was starting to think about how to solve distributed systems complexity. I rode that cloud-native mega-trend.
Today, intelligent agents and AI are an unstoppable mega-trend. Even though my compensation and environment at Kong were excellent — I was a founding veteran, many people told me I could retire — I still chose to leave, because the wave was here.
I didn't see any sufficiently deep, battle-tested Infra team working in this direction. Recognizing the swell, having the courage to charge at the crest, and knowing when to paddle home when the tide recedes — this sensitivity to the times and this courage are indispensable.
How do newcomers pick winners among startups?
👦🏻 Koji
You were on the founding team at both Cloudflare and Kong, among the earliest to join. In hindsight, you did pick the industry winners. What's the secret to that eye?
👨🏻💻 Guanlan
This probably mirrors your investment thinking, Koji. You invest capital; I invest my youth and energy.
The secret still lies in people: looking at the team's caliber, whether they're the most daring and hardworking, the most technically grounded crew.
Once you have direction, execution depends on the team's ability to solve problems at the foundational level — whether they can "go one layer deeper."
The best engineers building large model Infra today can definitely drill down to the kernel level, solving operator scheduling and protocol-layer performance. Back when I was hiring gateway leads at Kong, I directly tested how deeply they understood the network protocol stack.
90% of people think this stuff is good enough as-is. Only 5% will drill down to study underlying principles out of intense curiosity.
I love working with this tiny minority of geniuses who can rapidly digest and hand-build systems — this is also my standard for picking companies.
Young people have limited information — how do you truly understand a startup before the interview?
👦🏻 Koji
Young job seekers have very limited access to company information, often just brief exchanges during interviews. How did you quickly evaluate and understand Cloudflare and Kong back then? Any techniques or experiences worth sharing?
👨🏻💻 Guanlan
Young people's most abundant and precious asset is time. You can spend more of it, even run a thought experiment: if I were their CEO, facing current competition, what strategy would I choose? When you've done this kind of deep strategic推演, then go talk to them — through aligning cognition and perspective, you can easily see their true level.
Also, do your homework. Before joining Kong, I had on-and-off conversations with the founders for over a year. I only made my decision when their strategic pivot had matured in those final months.
Listen to more investment podcasts like yours, Koji. Evaluating a career and evaluating an investment share identical first principles. When your cognition and their business reach perfect match, your inner voice will signal: not joining would be a lifetime regret.
$3 million for angel investing — who gets it?
👦🏻 Koji
If you had $3 million in virtual capital, required to deploy it today, who would you invest in?
👨🏻💻 Guanlan
I won't name specific names. I can give two directions:
One is early founding members of top-tier video model teams. I've been helping them fundraise recently, and will even participate personally. Video models are the second domain, after coding, that's been proven to have real monetization potential, but there are very few startups that can actually do it well.
The other is teams in the AI for AI space. Looking at how to more intelligently train and improve large models through Agent long-horizon invocation and Auto Research. I have my selfish interest too — I want to work with such frontier teams to jointly refine our execution substrate and define the next paradigm of compute.
Claude Code or Codex?
👦🏻 Koji
Do you prefer Claude Code or Codex?
👨🏻💻 Guanlan
I use Codex now.
👦🏻 Koji
What's the reason? Everyone probably has a migration story.
👨🏻💻 Guanlan
Codex is all business, no fluff.
I remember someone sharing: Codex is more like a German — quietly grinding away, high quality; Claude Code is more like an American — very chatty, lots of filler, constantly checking if you've thought it through.
Also, Codex doesn't randomly ban accounts. Our US-registered accounts get banned frequently with the other one, it's a hassle to manage, have to apply for unbanning. Even though I know people at Anthropic, constantly reaching out gets depressing.
Codex frequently offers resets, developer-friendly. Its harness is also open-source. It pulled us over from Claude Code to Codex. In this coding wave, it's established itself, worth running and using.
Although when running on our Execution layer, we've found Codex has issues too. But because it's open-source, we plan to contribute fixes upstream. We want to optimize the harness above through our execution substrate.
👦🏻 Koji
Is everyone internally standardized on Codex? Or are people still using Claude Code, Cursor?
👨🏻💻 Guanlan
I personally use Codex quite a bit, also Grok and some domestic frontier models on Fireworks.
Our internal consensus is to completely leave Model Router choice to engineers, because they have the best intuition.
Our internal usage is all over the map. For frontend development or daily琐事 with larger fault tolerance, we'll use Grok or domestic frontier models on Fireworks.
I don't care what models people use — I just need them to be accountable for the final output. Through this freedom of choice, we cultivate their intuition and sensitivity to frontier model evolution.
👦🏻 Koji
Besides Codex, what AI products do you personally use most?
👨🏻💻 Guanlan
I also personally use Hermes and Pi. Our team is more interested in hand-rolling Harness — we often end up building something more usable than the open-source alternatives — so Harness itself doesn't have a deep moat. Eventually every company will choose to build their own.
I also use ElevenLabs for voice recording, and I experiment with all kinds of image generation models. I'll try out novel agents too — like a smart voice agent that helps you get refunded for overcharged phone bills from carriers. I'm probably the most aggressive adopter of agents internally.
Entrepreneurship: Five Good News, Five Bad News Every Day
Koji
I feel like you're in a really energetic, happy state overall.
Guanlan
Being a startup CEO really is like a roller coaster. You might encounter five pieces of good news and five pieces of bad news every day. The anxiety comes from always focusing on the bad news.
But the core question is: are we solving a problem exciting enough? Behind Runta, we have Jeff Dean, Fei-Fei, Ali Ghodsi, and the founders of Hugging Face — the top systems researchers and business leaders — supporting us.
They genuinely believe we're building foundational infrastructure that can profoundly reshape the future industrial landscape of AI and agents. That sense of extreme meaningfulness gives me endless excitement.
Koji
That's wonderful, that state is enviable. Thank you so much, Guanlan! I hope your joy and entrepreneurial passion can inspire the listeners of this episode.
Guanlan
We hope that five years from now, people will no longer debate whether "agents need a dedicated execution substrate" — just as today no one debates whether data in databases should be persisted, or whether internet services need identity and access control.
When billions of agents truly start working, they'll need their own compute environments, permission boundaries, audit systems, and recovery mechanisms.
If this thesis holds, what we're building now isn't an agent tool — we're helping define the next generation of computing infrastructure.
That's what Runta wants to do.
If you're an engineer who loves kernels, distributed systems, virtualization, and agents, welcome to join Runta. (Email: hiring@runta.com)
Koji
Thank you so much, Guanlan. Let's wrap up here for today, thanks!
Guanlan
Great, thanks Koji, bye!
