Decagon, a 24/7 AI customer service agent platform, scaled from 50 to 500 employees in one year while developing a sophisticated forward deployment engineering model. The company addresses the challenge of deploying agentic AI systems across diverse enterprises and verticals, from Fortune 20 companies to mid-market brands. Their solution involves splitting forward deployment into two specialized roles: agent builders who configure AI agents using natural language within the UI, and agent software engineers who translate enterprise feature requests into scalable product features. The results include successful deployments across multiple verticals including financial services and tech companies, with a focus on rapid value demonstration, knowledge compounding across deployments, and systematically converting custom work into self-serve features that scale across the entire customer base.
Decagon provides a 24/7 AI customer service agent that replaces traditional IVR systems and long email response times with human-like AI interactions across multiple channels. The platform is multilingual and omni-channel, initially landing with customers to handle complex support workflows that traditionally require human agents, then expanding into revenue-generating activities. A notable example is Hertz, which started using Decagon for inbound support deflection but expanded to proactive customer outreach for lease renewals and extensions. The company experienced hypergrowth from 50 to 500 employees over the course of a year, necessitating significant evolution in their LLMOps and forward deployment practices.
Decagon operates with two distinct types of forward deployment engineering activities. The first involves configuring the AI agent’s “brain” to work for specific enterprise contexts, similar to training a human employee. This includes defining instructions for handling user queries, establishing brand tonality and voice, and configuring actions the agent can take on behalf of users. The second type focuses on capturing enterprise product requests from the field and ensuring these get incorporated back into the core product in a scalable way. A critical insight shared is that forward deployment engineers at Decagon serve as the frontline for identifying patterns across enterprises, with the responsibility to recognize when enterprise A’s request will soon be needed by enterprises B, C, D, and E.
A fundamental organizational principle at Decagon is that forward deployment engineering is identical to product engineering in terms of bar, reporting structure, and often team composition. This reflects the blurred line between traditional forward deployment work and product engineering when working with Fortune 20 companies, where enterprise pain points frequently translate directly into product features that need to be built and prioritized. This structural decision ensures that field insights are systematically incorporated into product development rather than being siloed in customer-specific implementations.
In the early days at 50 people, Decagon employed agent software engineers who handled all aspects of deployment: configuring the agent brain alongside customers, integrating with backend systems for actions like password resets and more complex operations, and building platform features based on customer requests. As the company scaled to 500 people, this role was deliberately split into two specialized lanes to enable scaling.
The first specialized role is the agent builder, described as “Decagon pros” with deep intuition for the various models powering the platform. These individuals understand how to make the models work for specific enterprise use cases, largely operating within the UI to the extent possible. They flag when capabilities need to move beyond the UI and help bring those needs into the product roadmap. The second specialized role remains agent software engineers, who serve as the frontline for enterprise product requests and ensure these get incorporated back into the product systematically.
A particularly insightful theme emphasized throughout is the concept of exercising restraint in an era where AI coding tools make it temptingly easy to quickly implement one-off solutions. The speaker notes that with AI coding tools like Codex and Claude Code being so capable, the scarce skill is actually restraint—the discipline to think carefully about how a solution will scale to future customers rather than simply prompting an AI to generate a quick fix. This is especially important given Decagon’s ethos of building agents to be owned by the customer; black box solutions created from prompts and patches become too brittle and don’t serve either party well.
This principle extends to the broader requirement gathering phase. As AI coding has changed engineering generally, more effort must go upfront into requirements gathering and ensuring alignment on what needs to be built before implementation begins. This is particularly true when dealing with large enterprises where there’s a temptation to just get started quickly.
Decagon has developed practices to systematically scale deployment knowledge across their organization. One key approach is staffing deals with industry experts who have experience in the relevant vertical. For example, if working with financial services companies A, B, and C, when financial service company D comes along, the same core group with financial services experience gets assigned. This enables several benefits: the ability to speak in industry-specific terminology which builds credibility, faster ramp-up times, and compounding knowledge about how to configure agents successfully for that vertical.
The overarching goal is making every deployment faster than the last one, achieved through systematic knowledge transfer and pattern recognition across similar enterprises. This approach acknowledges that different verticals and company sizes require vastly different forward deployment approaches, making domain expertise particularly valuable.
A central principle in Decagon’s LLMOps philosophy is the mantra “custom becomes self-serve.” Whenever someone at Decagon has to do something manually, they work to upstream it back into the product as a self-serve capability. The integration story provides a concrete example: early in the company’s history, they built custom integrations to various CRMs repeatedly. After approximately the 25th custom integration, they recognized the pattern and built a self-serve integration capability. What previously required custom engineering code can now be self-served by customers or built by the agent building team.
This principle reflects a deliberate funnel model where forward deployment engineering serves as the frontline for customer asks, but the goal is always to scale these solutions across the business and ultimately make them self-serve. The company’s ethos holds that customers should be able to configure the agent completely via natural language, and if an engineer needs to intervene, that represents something that should be upstreamed into the product.
Particularly when working with Fortune 500 enterprises, Decagon prioritizes proving value as quickly as possible rather than pursuing multi-month timelines to demonstrate impact. While Decagon’s platform can become arbitrarily complex with support across numerous channels and user intents, the deployment strategy focuses on demonstrating value immediately with a narrower scope, then expanding once value is established. This recognizes that customer relationships are multi-year partnerships, but initial trust and momentum require fast wins.
This approach requires careful scoping at deal inception—literally during the very first conversations—to define what success looks like for the customer. This includes getting specific about metrics, channels needed, and desired outcomes, ideally in writing to prevent miscommunication. The upfront investment in precise requirements enables the team to race toward delivering defined value rather than pursuing an ambiguous or overly ambitious initial scope.
Forward deployment personnel at Decagon are positioned as advisors, not just executors of customer requests. Customers often arrive with specific requests about what they want automated or implemented, and while they’re frequently right, the forward deployment team has unique knowledge from seeing patterns repeated across every customer. Decagon ingests customers’ historical support data and uses this to advise on which workflows to automate first for highest ROI, which sometimes differs from what the customer initially requested.
This advisory positioning recognizes that forward deployment teams accumulate domain expertise that’s extremely valuable for customers to tap into. They see across many companies and can bring best practices and data-driven insights that individual customers cannot access on their own. The role encompasses both execution of customer requirements and strategic consultation based on cross-enterprise learnings.
As Decagon grew from 50 to 500 people, thinking shifted significantly toward designing systems for scale rather than purely optimizing individual deployments. The company has become rigorous about sharing knowledge across deployments, ensuring that information from the field feeds back to the platform systematically. The principle is that every time the agent interfaces with one customer, improvements should benefit subsequent customers—the agent should compound with each enterprise interaction.
This systematic approach to capturing and redistributing field learnings represents mature LLMOps practice, where production deployments become a source of continuous improvement for the platform itself. The knowledge flow from field to platform becomes formalized rather than ad hoc, ensuring that the benefits of enterprise scale are actually realized through compounding improvements.
Three key factors are identified as driving Decagon’s success in the market. First is a reputation for moving extremely fast on customer requests, attributed to a hard-working team culture. Second is earning trust as advisors rather than just executors, being able to provide guidance based on cross-customer insights and data. Third is effectiveness at productizing custom work, though the approach to this has evolved significantly as the company scaled from fitting around a lunch table to 500 people. The focus now is on designing systems that enable knowledge sharing across deployments and ensuring field feedback systematically improves the platform for all future customers.
The production LLM system must handle significant complexity across dimensions. Decagon operates across multiple communication channels including phone calls, email, text, and WhatsApp, requiring the AI agent to maintain consistent capabilities and personality across these different modalities. The system also spans multiple verticals from financial institutions to tech companies, with notable customers across various industries. The deployment approach must account for vastly different requirements based on both enterprise size and vertical, necessitating the industry expertise approach to staffing engagements.
A core technical philosophy is that agent configuration should happen entirely through natural language rather than requiring code-level intervention. This represents an important LLMOps principle where the interface for customizing AI behavior is itself natural language, making the system accessible to non-engineers and reducing the deployment friction. When engineers need to intervene in ways that go beyond natural language configuration, this signals a gap in the platform that should be addressed by expanding self-serve capabilities. This philosophy drives continuous platform evolution toward greater accessibility while maintaining the sophistication needed for complex enterprise use cases.
The AI agents require extensive integration with customer backend systems to take actions on behalf of users. These integrations range from relatively simple operations like password resets requiring authentication system access, to far more complex workflows. The evolution from custom-built integrations to self-serve integration capabilities represents a maturation of the platform’s LLMOps infrastructure, reducing deployment time and engineering overhead while expanding what customers and agent builders can accomplish independently. This pattern of identifying repetitive integration work and converting it to reusable, self-serve capabilities exemplifies effective scaling of an LLMOps practice.
Vercel addresses the challenge that while AI models have democratized the building of agents and internal tools, production deployment at scale remains difficult. The company built d0, an internal analytics agent that answers hundreds of data questions daily, using their own agent orchestration platform. By leveraging Vercel's infrastructure primitives—Sandboxes for isolated execution, Fluid Compute for dynamic scaling, AI Gateway for multi-model routing, Workflows for durable orchestration, and built-in observability—one engineer built d0 in weeks using only 20% of their time. The platform now supports multiple internal agents (lead qualification, customer support handling 87% of initial questions, abuse detection, content generation) and customer-facing products (v0 code generation and Vercel Agent for PR reviews), demonstrating how purpose-built infrastructure enables rapid development and reliable operation of AI agents without requiring deep DevOps expertise.
Anthropic developed a production multi-agent system for their Claude Research feature that uses multiple specialized AI agents working in parallel to conduct complex research tasks across web and enterprise sources. The system employs an orchestrator-worker architecture where a lead agent coordinates and delegates to specialized subagents that operate simultaneously, achieving 90.2% performance improvement over single-agent systems on internal evaluations. The implementation required sophisticated prompt engineering, robust evaluation frameworks, and careful production engineering to handle the stateful, non-deterministic nature of multi-agent interactions at scale.
Exa.ai has built the first search engine specifically designed for AI agents rather than human users, addressing the fundamental problem that existing search engines like Google are optimized for consumer clicks and keyword-based queries rather than semantic understanding and agent workflows. The company trained its own models, built its own index, and invested heavily in compute infrastructure (including purchasing their own GPU cluster) to enable meaning-based search that returns raw, primary data sources rather than listicles or summaries. Their solution includes both an API for developers building AI applications and an agentic search tool called Websites that can find and enrich complex, multi-criteria queries. The results include serving hundreds of millions of queries across use cases like sales intelligence, recruiting, market research, and research paper discovery, with 95% inbound growth and expanding from 7 to 28+ employees within a year.
