Stacks/Customer support

Decagon's fine-tuned specialist agent models

Decagon's research post describes a production customer-agent architecture built from many fine-tuned specialist models — intent, workflow selection, hallucination detection, escalation — rather than one frontier model. Smaller models are specialized with SFT and then RL (GRPO, GSPO, or DPO) so they can meet voice-latency constraints while outperforming larger general models on their narrow tasks. The stack is explicitly model-agnostic across providers.

Use case

Customer-facing conversational and voice agents, where Decagon uses a network of specialized models rather than one monolithic model, including for real-time voice where latency matters

Category

Customer support

Team size

not stated

Orchestrator / framework

Architecture is a network of specialized models in coordination, each with a distinct function: interpreting user intent, executing workflows, detecting hallucinations, or deciding when to escalate to a human. Workflow selection is framed as choosing the right Agent Operating Procedure (AOP) for a user's query. The authors say this orchestration is what a single large model can't guarantee on its own.

Memory

not stated

Evals

Fine-tuning is described as a continuous process of evaluation and refinement from real-world deployments. Training uses supervised fine-tuning on curated input-output pairs (for example, historical queries paired with the correct AOP) and reinforcement learning with techniques named as GRPO, GSPO, or DPO, typically SFT first and then RL. Training data is cleaned and filtered to remove PII and protect customer IP.

Observability

not stated

Monthly cost

not stated

Models

Fine-tuned smaller specialist models (providers not named; the team says it is model-agnostic and experiments across multiple providers)

Observability tools

not stated

Other tools

Agent Operating Procedures (AOPs)Human escalation

Source

https://decagon.ai/blog/fine-tuning-ai-agents

Max Lu · Nov 19, 2025