AI Support Agent Platform

8/6/2026

This post details how Decagon's conversational design team reduced 'barge' rates in billing dispute calls by over 15% through specific conversational design changes. These changes include reducing response length to prioritize critical information, eliminating system-serving questions in favor of those that advance the caller's goal, and implementing conversational state management to prevent repetition of bill details. These insights have been codified into guidelines for billing workflows and are actionable via Decagon's Agent Operating Procedures (AOPs) and Duet platform, enabling rapid iteration and deployment of conversational improvements.

7/9/2026

This post details the engineering effort required to deploy Decagon's conversational AI platform within a customer's VPC for air-gapped environments. Key technical contributions include building a self-hosted telephony layer from scratch using SIP signaling and media handling primitives, enabling self-hosted model inference for Decagon's proprietary models, and developing custom connectors for deep integration with customer-specific internal tooling, authentication systems, and data systems. The post also outlines different deployment models (SaaS, Single-tenant SaaS, Cloud-prem/VPC) and their associated engineering trade-offs.

6/30/2026

Introduced the next generation of Decagon's Simulations testing suite for AI agents. Key technical advancements include: auto-generating tests from AOPs and production conversations using Duet to create a self-maintaining golden test set, automatic detection and updating/removal of outdated tests, structured conversational arcs with checkpoints for granular results, flexible assertions for evaluating agent behavior across multiple dimensions, and native integration into CI/CD pipelines for production-grade testing.

6/9/2026

Introduced DuetBench, a new evaluation framework for self-improving customer service agents. DuetBench measures the agent's ability to diagnose issues, build Agent Operating Procedures (AOPs), generate and validate test simulations, and perform end-to-end certification tasks. The framework mirrors real-world enterprise agent building with simulated customer deployments, AOPs, tools, guardrails, and synthetic traffic. Results show Duet Autopilot outperforms human baselines in diagnostic tasks (93% vs 83%), timed agent-building tasks (45.5% vs 23%), and simulation accuracy after self-critique (58% to 88%). Autopilot also achieved the human certification standard (90% pass rate). This demonstrates a shift towards agent managers overseeing automated agent-building work.

5/28/2026

Introduces QA Hub, a new component of the AI Support Agent Platform, designed to enhance human-in-the-loop quality review. QA Hub provides a flexible workspace for organizing QA batches, applying custom rubrics, and tracking follow-ups, with features for team collaboration and external sharing. It integrates with Watchtower for automated QA and leverages Duet for synthesizing feedback and identifying improvement areas. This post details the workflow and benefits of QA Hub for improving AI agent quality through a combination of human expertise and automated analysis.

5/14/2026

Introduces 'Guided Discovery', a new capability for AI Support Agents that enhances their ability to handle exploratory conversations. This is powered by next-generation Agent Operating Procedures (AOPs) which allow agents to work from higher-level, natural language instructions, maintain full context across users and sessions, and reason toward a goal rather than follow a fixed path. This enables more adaptive customer experiences in product discovery, retention, and expansion scenarios.

4/14/2026

This post details how Decagon enhances the AI Support Agent Platform's reliability by addressing the limitations of using Model Context Protocol (MCP) alone. It introduces architectural improvements for tool curation and evaluation, including schema refinement to create narrower, intent-scoped tool definitions, layering guardrails on MCP definitions, and implementing routing logic via Agent Operating Procedures (AOPs) to constrain the tool surface area presented to the agent. Furthermore, it emphasizes the importance of ongoing evaluation at the tool level to catch regressions, ensuring systematic improvement and reliability in production.

4/2/2026

This post details the architectural decision to build Decagon Duet on Agent Operating Procedures (AOPs) rather than connecting a frontier reasoning model directly. It explains how AOPs provide a structured framework for Duet to analyze transcripts, suggest workflow updates, and take action within the Decagon platform. The post highlights the benefits of this approach, including direct access to platform tooling, improved visibility for non-engineering teams, and the ability for Duet to self-improve its own AOPs.

3/25/2026

This post details the optimization of the reranking stage within the RAG pipeline for low-latency AI agents. It explains the difference between list-wise and point-wise rerankers, highlighting the latency benefits of point-wise reranking due to shorter sequence lengths and predictable per-call costs. The core technical contribution is the implementation of efficient batching using attention masking, which allows multiple (query, document) pairs to be scored in a single batched forward pass by creating a block-diagonal attention pattern. The post also discusses finding the optimal batch size for minimizing end-to-end latency, empirically showing batch size 2 as optimal for their system.

3/25/2026

This post details the limitations of Supervised Fine-Tuning (SFT) for AI agent conversation classification, specifically the issues of compounding errors at inference, the ineffectiveness of self-distillation, and the erosion of steerability. It introduces 'steerability' as a critical metric, measured by a benchmark that quantifies a model's response to perturbations in category definitions. The post then discusses on-policy distillation as an improvement over SFT by closing the train-inference gap, but notes its dependency on teacher quality. Finally, it advocates for Reinforcement Learning (RL) with verifiable rewards as the most robust approach, enabling models to generalize beyond teacher supervision and discover correct reasoning strategies. The optimal training progression is presented as SFT -> on-policy distillation -> RL, each stage addressing the shortcomings of the previous one to create a more robust and correctable AI agent.

3/19/2026

Introduces Duet, an AI partner for agent building. Duet automates AOP generation from customer transcripts, creates tailored test scenarios, and analyzes production conversations for optimization. It aims to compress agent development cycles by enabling non-technical users to build and refine agents, moving beyond manual iteration and engineering dependencies.

3/9/2026

Introduces Agent Workbench, an AI chat assistant integrated into the AI Support Agent Platform. Agent Workbench provides consolidated access to debugging information such as audit logs, agent reasoning, and tool errors, enabling users to identify root causes, understand AI behavior, and receive actionable guidance for workflow refinement and issue resolution without direct engineering involvement.

3/9/2026

This post introduces Proactive Agents, which enhances the AI Support Agent Platform with user memory and outbound voice capabilities. User memory allows for persistent conversational context and structured metadata capture (e.g., feature requests, preferences) across interactions. Outbound voice enables proactive customer engagement via initiated calls, with built-in handling for voicemail, robocall screeners, and regulatory compliance. Agent Workbench is introduced as a self-serve debugging tool that leverages agent logs, reasoning, traces, and latency events to provide actionable insights for performance improvement.

11/21/2025

This post details Decagon's strategy for engineering fast and performant AI agents by fine-tuning specialized models. It explains the limitations of off-the-shelf models for customer-facing tasks and introduces the architecture's reliance on a network of specialized models. The post elaborates on two key fine-tuning techniques: supervised fine-tuning (SFT) for training on curated examples and reinforcement learning (RL) for optimizing based on quantifiable rewards. It highlights the benefits of this approach, including improved accuracy, reduced latency, and modularity for continuous evaluation and refinement, enabling model-agnostic experimentation and rapid performance evolution.

11/5/2025

Introduces Agent Versioning to the AI Support Agent Platform, enabling structured, data-backed evolution of AI agents. This includes isolated Workspaces for safe experimentation, Git-based tracking for versioning and validation, and enterprise-grade governance with structured reviews and auditability. It applies CI/CD principles to agent workflows, allowing for testing with unit tests or simulations before production deployment, and uses data from rollouts to compare metrics and benchmark new versions.

8/20/2025

This post details the application of Decagon's AI Support Agent Platform to the travel and hospitality sector. It highlights how AI agents are used to handle disruptions and seasonal surges, resolve bookings and account needs, act as personal travel concierges, convert service moments into revenue opportunities (upsells, save-the-sale incentives), and generate operational intelligence through conversational data analysis. Specific Decagon tools mentioned are Agent Operating Procedures (AOPs) for business rule adherence, guardrails for brand consistency and escalation, Watchtower for continuous QA, and Ask AI for querying historical interactions. The post emphasizes the platform's ability to integrate with existing systems (booking engines, CRMs, payment platforms) and operate across multiple channels (chat, email, voice).

8/6/2025

This post details how Decagon's AI agents are being used in the retail industry to resolve order and account-related questions, manage seasonal surges, help customers discover and buy products, deliver consistent on-brand support, and spot hidden friction. It highlights the use of Agent Operating Procedures (AOPs) for consistent resolution and tone, integration with platforms like Shopify, and the application of 'Ask AI' and 'Watchtower' tools for analyzing customer feedback and identifying trends. The post also outlines the future vision of AI agents delivering true concierge experiences by anticipating intent and surfacing upsell opportunities.

4/29/2025

This post introduces Decagon's AI Support Agent Platform and its core concept of Agent Operating Procedures (AOPs). It details the strategic considerations for building vs. buying AI support agents, highlighting Decagon's approach to customization, risk mitigation, and integration. The post also outlines how the platform addresses multi-channel challenges (voice, chat, email) and simplifies operational overhead and scalability through automated training, content management, monitoring, and robust infrastructure. It emphasizes a hybrid approach where Decagon's platform can be combined with internal enhancements.