Interview, Fireside Chat
How Decagon Runs 90% of Its Agents on Open-Source Models
- Strategic Narrative Shift (2026): The prevailing market narrative in the first half of 2026 posits that Frontier Labs (Anthropic, OpenAI) are the "last startups," with the belief that AI agents will replace all software applications. Decagon counters this by asserting that even with AGI, agents will still require dedicated software infrastructure to store work, reason, and interact with legacy systems.
- Open Source vs. Frontier Model Adoption:
- Current Ratio: 90% of Decagon's workflow now relies on open-source models, with the remaining 10% reserved for closed-source frontier models.
- Performance Trade-off: While frontier models offer general intelligence, Decagon found that fine-tuned smaller open-source models outperform state-of-the-art frontier models on specific, narrow tasks (e.g., latency-sensitive voice agents).
- Latency Optimization: The primary driver for adopting open-source models was reducing latency in voice interactions, a constraint frontier labs' small models could not meet without significant fine-tuning.
- Decagon Labs (Model Factory):
- Purpose: Established to compress the time between new model releases and producing task-specific, fine-tuned agents.
- Continuous Iteration: The landscape requires constantly training new models and deprecating old ones as capabilities shift; the team focuses on "completing the last mile" of deployment rather than just general model training.
- Evaluation Infrastructure: Decagon built proprietary evals and benchmarks specific to their use cases, as public benchmarks fail to measure customer outcomes effectively.
- Product Evolution & "Duet" Autopilot:
- Shift in Role: The company evolved from building a specific customer support agent to creating an agent that follows business processes generally.
- Duet Autopilot: A second, larger, and slower AI agent that automates the creation of "Agent Operating Procedures" (AOPs), writes integration code, generates tests, and monitors production conversations for trends and failures.
- Operational Impact: Duet replaced the manual engineering effort previously required to define procedures and tests, allowing non-engineers to deploy complex workflows via plain text instructions.
- Enterprise Deployment Strategy:
- Glass Box vs. Black Box: Unlike competitors like Sierra, which rely heavily on Forward Deployed Engineers (FDEs) acting as a "black box," Decagon utilizes a "glass box" approach where customers retain full visibility and control over the agent's logic and iterations.
- Speed of Iteration: A recent enterprise customer migrated from Sierra to Decagon, citing the ability to spin up seven new "journeys" (use cases) in one month compared to three in a year previously.
- Sales Cycle Acceleration: Decagon's sales-led approach includes pre-mapping enterprise model risk and compliance processes, allowing them to demonstrate a clear path from first meeting to full production deployment within 100% of the enterprise's constraints.
- Workforce & Hiring Dynamics:
- AI as Productivity Multiplier: Despite the "one-person unicorn" hypothesis, Decagon observes that AI tools lower the barrier to entry, prompting competitors to build faster, which necessitates hiring more talent to match the increased pace of innovation.
- Current Bottleneck: The primary constraint is no longer AI capability but talent acquisition; the company remains a "voracious consumer of tokens" but is constrained by the need for human decision-making on product direction and taste.
- Role of Forward Deployed Engineers: Initially necessary to discover workflows for new AI products, the long-term goal is to productize these learnings to eliminate the need for FDEs, transforming them into core product engineers.
- Market Trends & International Expansion:
- Global Pull: International expansion (e.g., Australia, London) is driven by top-down board pressure to adopt AI and reduced language barriers due to improved multilingual model capabilities.
- Data Residency: Global expansion requires navigating local data residency laws and local competitors, necessitating a "real investment" strategy rather than just remote support.
- Consolidation View: Decagon bets on horizontal SaaS consolidation (similar to Salesforce/Zendesk) rather than vertical isolation, believing scale and robust product depth win in the long term.
- AGI and Job Market Outlook:
- Jevons Paradox: Decagon observes that lowering the cost of customer support via AI does not result in layoffs; instead, it unlocks latent demand for service, leading to increased customer engagement and retention.
- Job vs. Career: The narrative is that AI will eliminate specific jobs (mundane, repetitive tasks) but not careers; human agents will shift to higher-value, revenue-generating interactions previously too expensive to offer at scale.
- SaaS Viability: SaaS and CRMs are not obsolete; they will evolve to serve as the "source of truth" and data storage for AI agents, rather than being replaced by them.
- Founder & Content Strategy:
- Social Media Dynamics: LinkedIn is used for product announcements and enterprise credibility, while X (Twitter) is leveraged to influence the "single timeline" of public sentiment and thought leadership.
- AI Personalization: Founders use custom-built agents to capture business context (hiring, deals, challenges) to enable AI to reason over their specific constraints before generating advice, moving beyond generic brainstorming.