Jul 30, 2026 · 1h 20m · a16z
How Decagon Runs 90% of Its Agents on Open-Source Models
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of The a16z Show, Decagon co-founders Jesse Zhang and Ashwin Sreenivas discuss how they built a leading enterprise AI agent platform by running 90% of production workloads on fine-tuned open-source models. They explore enterprise deployment strategies, productizing field observations, and why software and human careers will thrive in an AGI world.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The host as informed peer 5.0 Guest teaching 4.5 Guest disagreement 1.6 The host pushing back 1.5
speaking balance: gold is the host, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 21:49 Ashwin calls the tech Twitter FDE obsession a consulting trap
Ashwin forcefully dismisses the prevailing tech consensus that every AI startup should flood their ranks with forward-deployed engineers, warning that failing to productize field work reduces companies to glorified consulting shops.
Hardest push from the host ▶ 34:45 Co-host directly challenges Decagon's long-term right to exist post-AGIThe co-host refuses to accept standard SaaS talking points and directly presses Ashwin to articulate Decagon's actual 10-year moat if frontier model labs eventually achieve general intelligence.
Biggest teaching moment ▶ 4:59 Ashwin dismantles the model intelligence vs cost tradeoffAshwin corrects the widespread assumption that small open-source models sacrifice intelligence for price, proving that fine-tuned specialized models simultaneously win on accuracy, latency, and cost.
The host holds their own ▶ 1:17:36 Host connects expanded ticket volume directly to Jevons ParadoxThe host synthesizes Ashwin's real-world support ticket data into economic theory, precisely diagnosing the outcome as an empirical manifestation of Jevons Paradox in enterprise AI adoption.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The host as informed peer | Guest teaching | Guest disagreement | The host pushing back | Why |
|---|---|---|---|---|---|---|
| Welcome and Discussion Context on Open Source | 5 | 3 | 1 | 1 | The host opens by citing Jesse's viral post comparing open-source and frontier models to frame the discussion. Jesse walks through Decagon's transition from closed APIs to 90% open-source models to optimize voice agent latency. | |
| Why Fine-Tuned Models Outperform Frontier Models | 4 | 6 | 3 | 2 | Ashwin directly rejects the common Twitter narrative that choosing small models requires trading off intelligence for cost. He explains that task-specific fine-tuning allows smaller models to outperform frontier models on accuracy, speed, and cost simultaneously. | |
| Enterprise Adoption Hurdles for Open-Source AI | 4 | 5 | 1 | 1 | The host asks whether enterprise customers will follow Decagon's lead in fine-tuning open-source models internally. Jesse explains why enterprise inertia, custom evaluation requirements, and model risk governance will slow broad in-house adoption. | |
| Decagon Labs as an Enterprise Model Factory | 5 | 4 | 1 | 1 | The host asks for Decagon's framework on in-housing research talent versus buying ecosystem services like RL-as-a-service. Ashwin describes Decagon Labs as a continuous 'model factory' tightly coupled to end-to-end customer workflow outcomes. | |
| Tokenomics, Cost Priorities, and Company Stage | 4 | 5 | 2 | 2 | The co-host probes whether Decagon ignores model tokenomics compared to current industry hype. Jesse and Ashwin clarify that growth-stage companies prioritize conversational quality over cost, and open-source infrastructure naturally insulates them from token cost panic. | |
| AI Applications versus Infrastructure and Agent Labs | 6 | 4 | 1 | 2 | The host delivers an informed meta-critique against the prevailing narrative that frontier labs will eliminate vertical application startups. Jesse expands on why capturing nuanced enterprise business logic and QA compliance requires deep application software rather than raw model APIs. | |
| Industry Convergence and Software's Future Post-AGI | 5 | 6 | 4 | 2 | Ashwin counters the popular trend celebrating forward deployed engineers in AI, calling it a dangerous trap that devolves into glorified consulting if not converted to scalable product. The host eagerly invites him to elaborate on this contrarian stance. | |
| Evolving Forward Deployed Teams and Agent PMs | 5 | 4 | 1 | 1 | The co-host reflects on Decagon's early contrarian bet on Agent PMs and asks how the role has evolved. Ashwin explains that their deployment engineers and APMs focus strictly on translating client field pain points into universal product features. | |
| Product-Led vs Services-Led Enterprise Growth | 5 | 4 | 2 | 1 | The co-host asks Ashwin to contrast Decagon's approach with his previous deployment strategist role at Palantir. Jesse rejects the temptation to chase bespoke AI deployments, emphasizing that long-term enterprise software scale requires staying product-led. | |
| "Oh Shit" Moments, Duet Agent, and Post-AGI Careers | 5 | 5 | 3 | 2 | The host shares an anecdote about colleagues worrying AGI will eliminate careers and asks for concrete breakthrough product moments. Jesse playfully pushes back on post-work doom by arguing most modern jobs are abstracted human coordination, introducing the Duet meta-agent. | |
| Productizing Field Observations into Agent Procedures | 6 | 5 | 2 | 4 | The co-host poses a blunt question challenging Decagon's long-term right to exist if frontier models achieve AGI. Ashwin explains that their durable moat lies in enterprise software scaffolding, permissions, and legacy system interoperability rather than raw intelligence. | |
| Winning Enterprise Deals: Decagon versus Sierra | 6 | 4 | 2 | 1 | The host highlights Decagon's rapid emergence in enterprise deals against heavyweight competitor Sierra. Jesse shares a case where an enterprise switched from Sierra's services-heavy black box to Decagon's productized 'glass box' platform. | |
| Founder-Led Sales and Building GTM Velocity | 5 | 3 | 1 | 1 | The host questions how two technical founders cultivated elite commercial execution so quickly. Jesse explains their early hiring strategy of recruiting high-drive profiles and instilling high-velocity GTM DNA across the organization. | |
| De-Risking Enterprise Rollouts and Risk Processes | 6 | 4 | 1 | 1 | Ashwin details how mapping out model risk governance and staged rollout frameworks removes hesitation for regulated enterprise buyers. The host validates this from discussions with enterprise CIOs betting on founder execution velocity. | |
| Navigating Large Enterprise Orgs and Feedback Loops | 4 | 4 | 1 | 1 | Jesse shares that he spends roughly 80% of his time driving enterprise sales and breaking complex enterprise rollouts into piecemeal wins. Ashwin emphasizes the critical importance of keeping short feedback loops with live customer objections. | |
| Expanding Beyond Customer Support to AI Concierge | 5 | 4 | 1 | 1 | The co-host asks how Decagon expanded from narrow support tickets to an overarching AI concierge. Ashwin explains that their platform was architected to follow arbitrary business processes, unlocking inbound sales qualification and operational outreach. | |
| AI Product Roadmapping and Front Door Vision | 5 | 5 | 2 | 2 | When the host asks whether persistent memory or model capabilities are their biggest operational bottlenecks, Ashwin surprises her by stating that human hiring is the real constraint. Jesse outlines their long-term vision of agents serving as the comprehensive front door to businesses. | |
| Why AI Startups Keep Hiring & The Jevons Paradox of Engineering | 5 | 6 | 2 | 2 | The co-host inquires if AI productivity gains enable solo-founder unicorn setups. Ashwin counters using AI coding startups as an example of Jevons paradox, explaining that accelerated coding speeds cause competitive teams to expand engineering headcount and build three times as much. | |
| Unpacking Company Culture and Grind Slop | 5 | 3 | 2 | 1 | The host asks about online discussions surrounding 'grind slop' and Decagon's intense in-office culture. Jesse clarifies that their work ethic stems from genuine passion for building rather than performative posturing, while Ashwin highlights tight cross-functional collaboration. | |
| Preserving Culture Across Global Offices | 5 | 4 | 1 | 1 | The hosts ask how Decagon maintains cultural cohesion and operational quality across international hubs like London and Australia. Ashwin describes their immersion playbook of flying new hires to San Francisco and temporarily embedding veteran team members in new satellite offices. | |
| Horizontal Platforms vs Vertical Consolidation | 5 | 5 | 2 | 1 | The co-host asks whether niche local competitors can defend regional markets. Jesse argues that customer interaction software inevitably consolidates into horizontal platforms like Salesforce and Zendesk due to scale economies. | |
| The Role of CRMs and SaaS Survival | 4 | 4 | 1 | 1 | Jesse and Ashwin explain why CRMs will thrive as back-end systems of record for AI agents rather than disappearing. Ashwin shares a personal internal agent project designed to maintain persistent executive business context for decision-making. | |
| Social Media Strategy: LinkedIn vs. X | 5 | 5 | 1 | 1 | The host asks about founder brand building and distribution strategy between LinkedIn and X. Jesse breaks down how LinkedIn serves enterprise lead generation while X acts as the singular public timeline shaping downstream media and narrative reach. | |
| AI's Impact on Jobs, Jevons Paradox, and Career Up-leveling | 6 | 5 | 1 | 2 | The host probes the sensitive issue of AI labor displacement. Ashwin shows how lowering support costs dramatically expands customer demand for service, prompting the host to identify it as a classic demonstration of Jevons paradox, before Jesse concludes that AI eliminates mundane tasks while up-leveling careers. |
Statements from this episode (42)
Zhang: Fine-tuned small models can match or beat large frontier models
“And if you fine tune it to be really good at that task, it can be just as good or better than the big models.”
Zhang: 90% of Decagon's workflow runs on open-source models
“So today, 90% of our workflow is on open source.”
Sreenivas: Frontier Models Are Best Suited for Open-Ended Auxiliary Tasks
“So we think for jobs like that, Frontier models that are very smart, that can try out a lot of things, make a lot of sense.”
Zhang: Enterprise adoption of open-source AI will take longer than expected
“I think they'll get there, but it'll probably take longer than people think because, you know, even with our team, fine tuning these models is non-trivial.”
Zhang: Open-source AI is strictly better once production use cases solidify
“But I think the point is that at a certain point, it's strictly better to use open source models because when your use case is solidified and you're in production at scale and you're pretty sure this is the sort of shape of the agent, then there's no reason no…”
Zhang: Enterprise share of open-source AI inference is currently declining
“In enterprises right now, even though there's a lot of hype for open source, the sort of share of open source inference is actually Going down right now because people are spinning up all these new use cases.”
Sreenivas: Enterprise AI requires continuous model retraining rather than one-time projects
“We don't just build our set of open source models and then, you know, it's done, we can move on to our next thing, and maybe we'll revisit this in two years. You often need to train new models all the time because As the frontier changes, as the capability of …”
Sreenivas: Tailoring evaluations to customer outcomes beats tracking loss curves
“When we have open source open source models that we want to fine tune, we find that if we can clearly tailor our evals to customer outcomes, it's way better than just looking at, like, loss curves over time, right?”
Sreenivas: Optimizing AI agents breaks the cost, latency, and performance tradeoff
“Performance latency and accuracy is definitely the driving factor for most of this, right? Cost is a nice benefit in that, you know surprisingly, this is one of the few like tasks where you kind of get all the things for free, right? Like we don't actually hav…”
Zhang: Decagon's token usage per conversation increases to improve quality
“And actually over time, the number of tokens we're using per conversation has gone up because we're actually doing more model calls to make the quality better, to do more checks, to parallelize more things.”
Sreenivas: Deploying custom open-source models eliminates major AI cost pressures
“If we ran our entire business exclusively on frontier models, I would care a lot about costs, and I would think a lot about that, but once you're already, once you've already made the jump to saying, okay, now, now we know how to think about open source models…”
Zhang: Decagon fine-tunes models for use cases, not specific customers
“I think a common misconception that people have is, you know, fine tuning is, is a way to like customize it for that customer. In fact, most of the fine tuning we do is like customizing it for our use case, like the customer service use case.”
Zhang: Enterprise procedures must be taught to AI in-context, not fine-tuned
“I'm sort of teaching the AI my own procedures. And again, that doesn't happen through fine tuning. That, that happens like in context, because if you were to fine tune on that, you would have to reverse it every single time. You know, you change your procedure…”
Zhang: Vertical AI applications will outperform general lab-built agents
“The labs themselves will have more application capabilities, but those will be fairly general. Like they're, you can maybe build general agents that can do this thing or that thing. But for a lot of these like core verticals, like ours, our thesis is that, you…”
Sreenivas: AGI will not replace software because agents still need infrastructure
“I think even once you have AGI, all our AGI agents are going to need somewhere to store work and pull information from and reason about things. So I think, you know, a certain class of SaaS companies that were solely built for people to do work might face a bi…”
Sreenivas: AI application companies may eventually evolve into vertical model labs
“I think there will always be a space for application layer companies. Maybe in the longterm application layer companies just become labs for specific verticals, you know, because your primary product ends up being The models that are just really good at doing …”
Sreenivas: Early-stage AI startups need forward-deployed engineers for unproven workflows
“Forward deployed engineers are necessary or newly necessary for early stage AI companies because the workflows are new. Right? If you're building a SaaS company five years ago, most SaaS products are pretty well explored, right? Like you roughly know what the …”
Sreenivas: AI startups relying permanently on FDEs become glorified consultancies
“Once you know what the workflow is, you should not be relying on forward deployed engineers anymore, because once you know what the workflow is, If you can productize it, you should productize it and then become, you know, typical company with these scaling pr…”
Zhang: Very few AI companies can replicate Palantir's forward-deployed engineering model
“First of all, very few companies, if any, can do what Palantir does, which is like close massive deals off the bat. And like, it's kind of worth it to spend all that effort.”
Zhang: Human careers will still exist after AGI
“The first thing I want to say is I'm like, certain there will be careers after AGI. The reason for that is, like, most of our jobs, for sure are for jobs, are kind of, like, made up. Most jobs are made up... So when AGI is here, like it will change people's jo…”
Sreenivas: Current AI model capabilities far exceed what enterprises actually utilize
“The capability of models today is far greater than they are being used for within the enterprise, right?”
Sreenivas: Surrounding Infrastructure Will Be Primary Enterprise AI Need for Years
“And so I think for the next few years, that's probably going to be you know, the primary thing that these models need to be able to work.”
Zhang: Decagon's most recent customer churned from competitor Sierra
“Our most recent customer actually turned off of Sierra to come to Decagon”
Zhang: Selling enterprise AI requires pitching vendor approach over category demand
“Generally in these conversations, we're not really having to convince people to like, Invest in this space. It's, like, more of, hey, we're the right approach for you, so you can partner with us.”
Zhang: Decagon seeded early sales team with Ivy League athletes
“The other profile we had a lot of in the early days were just like Ivy league athletes, I guess. And those profiles were kind of a good foundation for the group.”
Zhang: Enterprise AI Rollouts Should Be Piecemealed, Not Deployed All at Once
“One of the things that we really try to do is, you know, kind of take the project and piecemeal it. So we're not just deploying across every surface area, every use case at once.”
Sreenivas: AI agents across support and sales fundamentally execute business processes
“The thing that we built and we kind of built this intentionally from the start was not an agent that does customer support well, but rather an agent that follows business process well. And executing on operational workflows, doing sales lead qualifications, an…”
Sreenivas: Improved model instruction-following unlocks open-ended workflows like sales discovery
“When you had models, you know, let's say a few years ago, you'd have to give it very, very tight guidance, very specific instructions that you didn't want it to deviate from, and as the models got smarter, you could kind of give it broader and broader guidance…”
Zhang: Rigid 12-month product roadmaps do not work in AI
“Realistically in, in today's AI world, it's very difficult to have, like, a 12 month roadmap to a T. You maybe know how, like, some themes of what you want to build, but ideally, if you have those things, you should just build it, like, right now, because it's…”
Zhang: AI agents should handle every customer interaction as business front doors
“An AI agent should just be the front door of your business, of your brand, and every interaction, whether it's like reactive or proactive with a customer should be handled by AI.”
Sreenivas: AI agents cannot yet decide what to build or exercise taste
“There are still things, I don't quite yet think we're at the point where we can have the AI agents make decisions on what to build, and kind of have that the taste of, is this done yet? Right? So we can outsource a lot of specific execution steps, but I don't …”
Sreenivas: AI coding startups use developer models heavily but still hire aggressively
“All the AI coding startups are hiring like crazy. You know, they're like the most sophisticated users, presumably, of these models, and they are hiring like crazy.”
Sreenivas: AI productivity gains will drive companies to build more, not fire
“Everybody has access to these tools, and so if our competitors are going to use them and build more things, we need to build more things, right? If somebody else said, oh, here's our roadmap, and now we can get through it in, you know, a third of the time, and…”
Sreenivas: Deep cross-functional teamwork requires employees to work in the office
“Being able to have teams that are so kind of disparate from like a function perspective all working together kind of one kind of needs people in the office because everybody's kind of jamming on ideas together”
Zhang: AI simplifies language localization, significantly accelerating software international expansion
“Language is a lot easier with AI. So in the past, maybe a blocker would be, oh, my language just doesn't work in, or my app just doesn't work in German or pick your language. But now it's a lot easier to adapt to your app. So I think for those reasons, interna…”
Zhang: Horizontal platforms will defeat pure vertical solutions in customer AI
“Our view, the reason why we've kind of built so horizontally is that we believe that in our space, the winners are going to be horizontal. There's just not that much that is like super verticalized that where like you could see a pure vertical solution survivi…”
Sreenivas: AI Agents Will Not Eliminate CRMs or Auxiliary Software
“Now, to the point of, Does that mean CRMs go away? I mean, my answer is no, because if you had you know, if you have a company today where your concierge is a human being, they still write your info on a CRM so that they can track it for later. So, and I think…”
Zhang: CRMs could thrive as source-of-truth databases for AI agents
“I actually think CRMs could do quite well. They'll, they'll be slightly different in the sense of, like, CRMs are kind of databases in a way, and, you know, the frustration people have with them sometimes is that the interfaces are not really difficult to use,…”
Sreenivas built a personal AI agent to gather context for executive decisions
“However, the bottleneck, I realize, at least for a lot of the work that I do, is business context, right? There's a lot of context for every idea around, okay, the constraints that we have, the goals that we're going for, and things like that that is difficult…”
Zhang: Solo founders reach conclusions much slower without peers to bounce ideas
“I do think like one of the big struggles being a solo founder is you don't have anyone to bounce ideas off of. So you just arrive at conclusions a lot slower.”