TechPrufer
Get started
Decagon
Decagon logoDecagon · Feature trail

AI Support Agent Platform

80 posts · Aug 2025 – Sep 2026

What this trail covers

Decagon is evolving its AI Support Agent Platform by introducing Duet Apprentice. This capability allows AI agents to learn from enterprise onboarding materials (Notion, Confluence, SharePoint, Google Drive) and observe escalated conversations handled by tenured employees. It also enables continuous learning by integrating with team messaging platforms like Slack and Microsoft Teams, proactively capturing changes and ambiguities. This aims to provide agents with a deeper understanding of business context and operational nuances, improving their autonomy and effectiveness.

All Decagon trails
  1. Autopilot in production: Governance of self-improving agents | DecagonThis post details the introduction of Duet Autopilot, a self-improving agent for agent building, and its supporting governance platform. Key technical contributions include: 1. Vault: A centralized, governed context layer for agents, supporting file structures and connectors to integrate existing documentation (e.g., Notion, Slack, SharePoint). 2. Outcome Orientation: Specifying which metrics (deflection rate, CSAT, AHT) should be prioritized when agent improvements compete. 3. Dual Autopilot Tracks: Trigger-based runs for immediate error response and Hillclimbing Autopilot for regular audits focused on deflection and CSAT. 4. Outcome-based Metrics: Comparing Autopilot-merged changes against baseline metrics (CSAT, deflection) before and after shipping, with projected impact on deflection. 5. Duetbench-2: A formal benchmark to measure the durability of accepted changes over time (e.g., 85% of changes remain effective after 30 days).
  2. DuetBench-2: Measuring agent self-improvement in production | DecagonIntroduced DuetBench-2, a more rigorous benchmark for evaluating AI agent self-improvement, focusing on verifiable capability, production impact, and compounding self-improvement. Detailed Autopilot's performance against this new benchmark, achieving an 81.9% pass rate for AOP building/revision, with specific metrics on build time and human intervention. Highlighted improvements in simulation generation (96% pass rate) and LLM-as-a-judge accuracy (99.5%). Quantified production impact with a 25pp resolution rate increase at a Fortune 50 company and demonstrated durability with over 85% of updates remaining live after 30 days. Provided an example of compounding self-improvement through a self-repaired polling loop in the simulation runner.
  3. Automated customer service without losing the human touch | DecagonThis post details how Decagon designs AI-powered customer service systems to balance automation with human interaction, focusing on resolution over deflection. It outlines principles for effective AI-to-human handoffs, including visible exits, context transfer, and channel continuity. The post also addresses preventing AI hallucinations through ground-truth enforcement, real-time monitoring, and gradual rollouts, and emphasizes measuring true resolution rate, re-contact rate, and abandonment by step as key metrics for evaluating automation success.
  4. Proactive customer service strategies for CX leaders | DecagonThis post details how Decagon's AI agents are being used to implement proactive customer service strategies. It outlines the process of identifying signals (e.g., stalled onboarding, repeating errors, slipping usage), designing triggers, and defining actions for AI agents. The post emphasizes the shift from reactive to proactive customer service, detailing how AI agents can take action within interactions (e.g., processing renewals, explaining benefits) rather than just sending notifications. It also introduces key metrics for measuring proactive programs, such as contact avoidance rate, deflection rate, outreach engagement rate, and signal-to-action time, and highlights the importance of A/B testing for establishing baselines. The post also touches on the operational changes for CX teams, moving from responders to architects of proactive outreach.
  5. Your Decagon agent can now search the live web | DecagonThis post introduces the integration of live web search into Decagon's AI Support Agent Platform, enabling agents to query the live web using Perplexity's Search API. This addresses the limitation of static knowledge bases by allowing agents to access and cite up-to-the-minute information. The technical contribution lies in the integration of an external search API to augment agent capabilities, allowing for real-time data retrieval to answer customer queries that were previously unresolvable or required escalation.
  7. Introducing Decagon Assist: Empower every representative to deliver concierge experiences | DecagonThis post introduces Decagon Assist, an AI copilot for human representatives, as an evolution of the AI Support Agent Platform. It focuses on providing human agents with automated summaries of prior AI agent interactions, consolidating necessary next steps into a single interface, and enabling direct querying of the organization's knowledge base. The system aims to integrate seamlessly across existing CRMs and channels, offering features like live transcription and real-time translation. A new analytics hub is introduced to measure adoption and performance impact.
  8. How conversational AI components work together in customer service | DecagonThis post details the architectural patterns and interface contracts necessary for reliably wiring conversational AI components together, emphasizing the importance of system design over individual component quality. It introduces concepts like response validation, exponential backoff, confidence thresholds, and state persistence to prevent cascade failures and ensure resilience. The operational framework includes knowledge ingestion via a Unified Knowledge Graph, natural language configuration through Agent Operating Procedures (AOPs), and continuous learning loops driven by AI observability. The post highlights the challenges of voice AI integration and the need for strict latency contracts.
  9. Getting closer to the build: How I helped create agent deployment engineering | DecagonThis post introduces the creation of the Agent Deployment Engineering (ADE) function. ADEs are customer engineers who translate customer requirements into production agents, focusing on technical build, testing, debugging, and iterative improvement. The role involves making tradeoffs between reliability, latency, precision, maintainability, and speed. The function aims to create reusable architecture patterns, tools, and testing practices, and to identify recurring platform and integration gaps to inform broader product development, fostering a compounding loop for agent building and product capability.
  10. "Agent, agent, agent": designing conversations to reduce barge | DecagonThis post details how conversational design improvements, specifically reducing 'barge' (callers demanding a human agent), can be achieved through iterative design passes. By analyzing real conversations and implementing changes to agent responses – such as saying less, leading with what matters, avoiding system-serving questions, and preventing repetition – Decagon has demonstrated a significant reduction in barge rates in billing dispute workflows. The platform's ability to deploy these natural language-based behavioral changes rapidly (within a day) is highlighted as a key enabler for such iterative improvements.
  11. What an air-gapped AI deployment actually requires | DecagonThis 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.
  12. Agent Operating Procedures: From Manual SOPs to Automated AI Logic | DecagonThis post introduces Agent Operating Procedures (AOPs) as a new capability within the AI Support Agent Platform. AOPs allow natural language instructions to be compiled into validated workflows for AI agents, enabling them to perform complex, multi-step tasks. The post details how AOPs differ from human SOPs, emphasizing explicit, machine-executable logic for AI. It highlights benefits such as faster implementation (3-6 weeks), increased omnichannel efficiency (70-80% deflection rates), and continuous improvement through simulation, unit tests, and regression testing. The process involves auditing existing procedures, translating them into AOP format with discrete steps, triggers, actions, and decision points, and connecting with existing tools like helpdesks and e-commerce systems. The post also briefly touches on alternatives in the AI customer service automation market.
  13. The next generation of Simulations: testing that keeps pace with your agents | DecagonIntroduced 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.
  14. Introducing Agent Development: How Decagon is redefining forward deployment | DecagonThis post details the evolution of Decagon's forward-deployed teams into a structured 'Agent Development' function. It highlights the shift from a 'hero' model to a 'systems design' approach for building AI agents. The core technical contribution is the reduction of custom engineering work per agent by 80% through the creation of a self-serve roadmap, enabling Agent Development Managers (ADMs) to handle build, debug, and testing in plain English, rather than requiring bespoke engineering sprints. This process aims to make agent builds more maintainable and customer-ownable, preventing them from becoming 'black boxes'.
  15. Knowledge base AI systems – how they really work | DecagonThis post details the technical underpinnings of AI knowledge bases, specifically how semantic search via vector embeddings, retrieval-augmented generation (RAG), and multi-source ingestion function. It outlines the four stages of RAG (embedding content, storing in vector DB, query matching, LLM generation) and the four-step automated content maintenance workflow (gap detection, stale content flagging, AI-assisted drafting, human review). It also highlights the importance of evaluating semantic search quality, automated content lifecycle, generative knowledge creation, analytics and gap detection, and integration depth when choosing a platform. The post positions Decagon's 'Suggestions' feature as an implementation of this automated maintenance workflow, drawing from conversations to identify gaps and generate drafts, and emphasizes integration with existing tools.
  16. What is real-time customer support and how to get it right | DecagonThis post details how Decagon's AI Support Agent Platform, specifically through Agent Operating Procedures (AOPs) and Watchtower, enables CX teams to manage AI behavior and monitor performance without engineering intervention. It highlights the shift from routing-focused tools to resolution-focused AI agents, emphasizing non-technical control for CX operators to update AI logic, escalation rules, and response guidelines. The post also discusses the economic impact of agentic AI in resolving customer service issues and reducing operational costs, citing examples like ClassPass achieving 24/7 support with significant cost reduction.
  17. Introducing Duet Autopilot: The self-improving agent for conversational AI | DecagonThis post introduces Duet Autopilot, a new capability for the AI Support Agent Platform. Autopilot automates the process of identifying issues in production conversations, tracing them to root causes, and proposing targeted fixes. It incorporates a self-validation loop using simulations against conversation data and curated test sets to ensure proposed updates are effective and do not introduce regressions. The system also enforces enterprise governance through human approval gates and provides detailed visibility into proposed changes and outcomes. The post highlights how Autopilot's own improvement loop leads to compounding gains in update quality over time.
  18. DuetBench: An evaluation of self-improving customer service agents | DecagonThis post introduces DuetBench, a novel benchmark for evaluating the self-improvement capabilities of AI customer service agents. DuetBench addresses limitations of existing benchmarks by focusing on dynamic, iterative work, the ability to self-improve, and real-world complexity. It comprises diagnostic, build, simulation-generation, and end-to-end certification tasks. The post presents results showing Duet Autopilot outperforms average human performance in diagnostic tasks (93% vs. 83% pass rate) and timed enterprise agent-building tasks (45.5% vs. 23% human pass rate within 3 hours), particularly for medium and hard complexity tasks. Autopilot also demonstrates strong self-critique capabilities, improving simulation accuracy from 58% to 88% and reaching the human certification standard (90% pass rate) for end-to-end agent building.
  19. QA Hub: Agent quality is a team sport | DecagonThis post introduces QA Hub, a new component of the AI Support Agent Platform, designed to streamline human-in-the-loop quality review for AI agents. QA Hub provides a dedicated workspace for organizing QA batches, applying custom rubrics, and tracking follow-ups, allowing CX teams to shape the workspace around their specific processes with configurable columns, filters, and saved views. It facilitates collaboration by enabling teams to organize batch reviews, assign owners, and tag teammates. The External QA Page extends this capability for external testers. QA Hub is positioned as complementary to Watchtower (automated QA), with QA Hub focusing on human expertise for building trust, pressure-testing rubrics, and incorporating diverse perspectives, while Watchtower provides broad coverage for known quality signals and failure patterns. Duet is integrated to assist in creating review batches, scoring conversations, and synthesizing reviewer feedback into actionable insights for agent improvement.
  20. Introducing Guided Discovery: Finding the right recommendation for every customer | DecagonThis post introduces Guided Discovery, a new capability for the AI Support Agent Platform. It details how next-generation Agent Operating Procedures (AOPs) are used to enable agents to handle open-ended conversations by reasoning toward a goal rather than following predefined paths. The post provides examples of its application in product discovery and customer retention, highlighting the use of user memory and adaptive tone.
  21. Contact center analytics and the KPIs that drive performance | DecagonThis post details the integration of contact center analytics into the AI Support Agent Platform. It outlines the four-stage analytics loop (Capture, Transcribe and Contextualize, Analyze, Activate) and key aspects like data collection, AI-powered analysis (sentiment, topic classification, keyword extraction, pattern recognition), and insight generation. It also describes common analytics types (speech, text, interaction, predictive, quality management) and their benefits, focusing on improving customer experience, supporting human agent performance, enhancing operational efficiency, and driving strategic decisions. Key metrics like AHT, FCR, CSAT, NPS, and call volume/hold times are discussed in the context of actionable insights. The post emphasizes the platform's ability to connect these analytics to Agent Operating Procedures for direct impact on agent behavior.
  22. Conversational IVR from definition to business impact | DecagonThis post details the technical implementation and business impact of conversational IVR, a new capability within the AI Support Agent Platform. It outlines the five-step sequence of conversational IVR (ASR, NLP/NLU, Dialogue Management, TTS, Continuous Learning), the intelligent routing mechanism with context transfer, and the importance of low latency. It also compares conversational IVR to traditional IVR, defines related terms like IVA and Conversational AI, and discusses business benefits such as higher containment and improved customer satisfaction. The implementation section highlights the importance of using real data, mapping conversational flows, configuring NLP/speech recognition, integrating with existing systems, and testing. Decagon's approach is described as a managed implementation where CX teams use AOPs for business logic while Decagon's engineers handle guardrails and integrations.
  23. The real difference between agentic AI and generative AI | DecagonThis post defines and differentiates generative AI from agentic AI, outlining the key characteristics of agentic systems: goal decomposition, persistent state and memory, and autonomous tool selection. It contrasts these with generative AI's content creation capabilities and discusses the limitations and risks of agentic AI in production, such as cascading errors and security exposure, advocating for robust guardrails and oversight.
  24. Understanding agentic workflows and the design patterns that power them | DecagonThis post elaborates on the technical underpinnings of agentic workflows, detailing design patterns such as planning (task decomposition), tool use, agentic RAG, reflection, and multi-agent systems. It categorizes agentic systems by autonomy and provides concrete examples of their application across various domains. The post also outlines key benefits (efficiency, scalability, reduced errors, faster problem-solving) and challenges (complexity, ethics, security, cost, error propagation, integration, compliance, reliability) associated with implementing agentic workflows. Best practices for implementation, including phased rollout, prompt engineering, and human-in-the-loop strategies, are also discussed.
  25. User memory: Persistent context that makes every interaction feel personal | DecagonThis post introduces User Memory as a new capability for the AI Support Agent Platform. User Memory is designed to store and reference conversational context and personal user signals across interactions, enabling personalized customer experiences. It captures unstructured conversation history and extracts structured metadata like feature requests and preferences. The system provides a unified user view, aggregates user analytics for trend spotting, and offers enterprise-grade governance with data portability, redaction, and expiration policies.
  26. AOP Copilot: Your AI assistant for building and optimizing AOPs | DecagonIntroduced AOP Copilot, an AI assistant that generates production-ready Agent Operating Procedures (AOPs) from natural language prompts. The system understands workflow step types and their connections to assemble structurally sound and validated AOPs, adhering to Decagon's best practices for clarity, modularity, and performance. It integrates directly into the Decagon UI for a streamlined drafting, revising, and deployment experience.
  27. Why MCP alone isn’t enough for reliable agent tool use | DecagonThis post details how Decagon addresses the reliability gap in agent tool use, which MCP alone does not solve. It introduces the concept of treating MCP as a tool registry and outlines Decagon's approach: schema refinement to split generic tools into intent-scoped definitions, layering guardrails on MCP definitions, and implementing routing logic via Agent Operating Procedures (AOPs) to constrain the tool surface area for agents. It also emphasizes the importance of ongoing evaluation with automated tests for tool selection and argument construction to prevent regressions.
  28. Why we built Decagon Duet on Agent Operating Procedures | DecagonThis post details how Decagon built its Duet AI partner on top of its existing Agent Operating Procedures (AOPs) framework. Instead of integrating a frontier reasoning model directly, Duet leverages AOPs to define its core processes (defining workflows, testing, refining logic, analyzing performance, diagnosing gaps) as executable procedures. This allows Duet to operate within the Decagon platform, access necessary tooling for actions like versioning and data analysis, and provides visibility to non-engineering teams. The post highlights that Duet can improve its own AOPs, leading to compounding self-improvement.
  29. Designing low-latency AI agents through reranker optimization | DecagonThis post details how Decagon reduced reranking latency in their RAG pipeline by transitioning from list-wise to point-wise reranking. They implemented an efficient batching strategy using attention masking to score multiple (query, document) pairs in a single batched forward pass, preserving scoring semantics while amortizing model call overhead. Empirical results showed that a batch size of 2 yielded the best latency improvements. This optimization was achieved through structural insight and targeted changes rather than overhauling the underlying model.
  30. Why off-policy training isn't enough: SFT, RL, and the limits of imitation | DecagonThis post details a novel training progression for AI classification models, moving beyond standard Supervised Fine-Tuning (SFT). It introduces a sequential approach combining SFT, on-policy distillation, and Reinforcement Learning (RL) to address limitations in model robustness and steerability. Key contributions include: 1) identifying SFT's structural limitations (compounding errors, steerability erosion), 2) proposing a steerability benchmark to quantify these failure modes, 3) demonstrating that on-policy distillation improves accuracy and steerability but introduces teacher dependency, and 4) showing that RL with verifiable rewards, initialized with SFT, leads to significantly more robust models that are correctable via prompt updates. The post also references prior work on fine-tuning AI agents and knowledge distillation.
  31. 3 Criteria for Agents | DecagonThis post outlines three key criteria for commercially viable AI agents: handling non-determinism with buffers and human-in-the-loop escalation, ensuring readily available data sources and tools for integration, and achieving repeatability and scalability across an organization. It emphasizes that successful agents deliver measurable value by addressing these challenges, drawing on Decagon's experience with enterprise customer support AI agents.
  32. The evaluation engine behind Decagon’s AI agents | DecagonThis post details Decagon's model evaluation engine for AI agents. It introduces a two-phased framework: offline evaluation using LLM-as-judge and ground truth benchmarks, and online A/B testing with real customers. The offline phase scores responses on relevance, correctness, naturalness, and empathy, with human audits for validation. The online phase tracks business-critical metrics like CSAT and resolution rate. The post emphasizes that evaluation extends beyond response generation to the entire agent architecture, including retrieval and reasoning, and highlights the continuous feedback loop for model and system improvement.
  33. Introducing Duet: Your agent building partner | DecagonThis post introduces Duet, an AI partner for agent building. Duet automates agent lifecycle management by identifying necessary workflows and generating AOPs, flagging and validating AOP updates, generating test scenarios, simulating conversations, and analyzing live conversations for tuning. It also scales improvements by analyzing production data to highlight trends, recurring issues, and workflow improvement opportunities, generating custom reports to guide efforts. Duet aims to compress the agent iteration loop from days to minutes.
  34. Agent Workbench: Debug AI agents autonomously | DecagonIntroduces 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.
  35. Introducing Proactive Agents: User Memory and Outbound Voice built for the AI Concierge | DecagonThis post introduces Proactive Agents, a new capability for the AI Support Agent Platform, by integrating user memory and outbound voice. User memory allows agents to retain conversational history and capture structured metadata (feature requests, sentiment, preferences) across interactions and channels. Outbound voice enables AI agents to initiate proactive calls, handling complexities like voicemail detection and regulatory compliance, powered by Agent Operating Procedures (AOPs). Agent Workbench is also launched as an AI chat assistant for debugging and optimizing agent performance by analyzing logs, reasoning, and latency events.
  36. The QA breakthrough: Watchtower turns every support interaction into actionable intelligence | DecagonThis post introduces Watchtower, an 'always-on QA tool' within Decagon's analytics suite, which significantly enhances the AI Support Agent Platform by providing comprehensive analysis of all customer interactions (voice, chat, email). Key technical contributions include: 1. Voice QA: Enabling complete conversation coverage for voice calls, tracking voice-specific metrics like transfer rates and tone, and allowing deep dives into underperforming conversations. 2. Rubrics: Introducing custom scorecards that go beyond basic 'issue resolved' metrics, allowing quantification and trend tracking of qualitative aspects like tone, relevance, and brand voice. This involves defining multiple fields with scoring options to measure conversational quality at scale. 3. Templates: Providing pre-configured, proven QA configurations for common categories (user sentiment, conversation quality, engagement opportunities) to accelerate insight generation and customization.
  37. Designing layered guardrails for reliable AI agents | DecagonThis post details the implementation of layered guardrails for AI agents to ensure reliability, safety, and brand alignment. It describes regression testing (single-turn and multi-turn interactions) before deployment, real-time guardrails during conversations (bad actor detection, human agent escalations, brand voice guidelines, response supervision) using separate fine-tuned models with minimal latency impact, and post-conversation analysis with Watchtower for continuous improvement. The technical depth includes details on how these guardrails function, such as parallel processing for bad actor detection and supervisor models evaluating responses for grounding and accuracy.
  38. Agent Operating Procedures: Decagon’s approach to AI agents for CX | DecagonThis post introduces Agent Operating Procedures (AOPs) as Decagon's product-centric approach to building AI agents for customer experience. AOPs combine natural language with code for dynamic response to complex issues and multi-step workflows, enabling non-technical users to shape AI logic while engineers maintain control. The system aims for faster development cycles (weeks vs. months), improved transparency into decision-making, and scalable, trusted results with enterprise-grade guardrails for security and validation.
  39. AI customer service automation: Ticket routing, response generation, and 24/7 support | DecagonThis post details Decagon's approach to AI customer service automation, focusing on context-aware automation for ticket routing, response generation, and 24/7 support. It addresses the 'speed-safety paradox' and 'reputation risk' by emphasizing action-oriented intelligence through Agent Operating Procedures (AOPs) and quality assurance layers like Watchtower and Guardrails. The post outlines principles for instant resolution, immediate escalation, and transparent handoffs, providing examples of automation candidates. It also details strategies for preventing AI hallucinations through a multi-model architecture and ground truth enforcement, and discusses the transformation of human agent roles. The implementation strategy involves a phased approach from pilot to platform, starting with specific, high-impact use cases.
  40. AI customer service challenges and solutions: A playbook | DecagonThis post contributes to the AI Support Agent Platform by detailing strategies to prevent chatbot failures, specifically hallucinations and knowledge base decay. It introduces architectural solutions like Retrieval-Augmented Generation (RAG) and action-constrained AI agents as primary defenses against incorrect information. It also highlights the importance of data privacy, compliance, and best practices for customer experience, including setting clear expectations, providing escape hatches, and using A/B testing. The post outlines a structured, collaborative setup process with Decagon and emphasizes a controlled failure approach with confidence thresholds and nightly evaluations as key to preventing failures.
  41. Climb higher: Introducing Diagnostic Tools for faster agent iteration | DecagonIntroduces Diagnostic Tools, a new feature within Decagon's Insights and Analytics suite, to accelerate AI agent iteration. This includes Escalation Drivers, which quantify conversation escalations by guardrail and AOP to prioritize refinements, and AOP Flow Analysis, which visualizes conversation progression through Agent Operating Procedures (AOPs) to identify workflow breakdown points. The system leverages Decagon's agent architecture with structured nodes for precise analytics and detailed workflow observability.
  42. The case for a unified, transparent platform in conversational AI | DecagonThis post introduces a unified platform for conversational AI, detailing the capabilities for building agents with Agent Operating Procedures (AOPs), optimizing performance through simulations, guardrails, and live experimentation, scaling intelligently with analytics and self-improving knowledge systems, and deploying across chat, voice, and email with shared memory and context. It emphasizes the integration of product, CX, and engineering teams around a single source of truth, moving away from fragmented internal builds or black-box platforms.
  43. Trace View: AI agents shouldn’t be black boxes | DecagonIntroduces Trace View, a new capability for the AI Support Agent Platform that provides granular visibility into agent reasoning and performance. Trace View offers a step-by-step interface detailing each interaction, including LLM calls, tool usage, and condition evaluations. It also provides latency metrics for each micro-step. This feature is a direct outcome of the Agent Operating Procedures (AOPs) architecture and enhances debugging, performance tuning, and validation through integration with unit testing.
  44. Understanding enterprise AI answering services | DecagonThis post details the technical and compliance considerations for enterprise AI answering services in production environments. It highlights Decagon's platform's capabilities in handling conversational complexity, achieving action-taking capability through deep system integration, and ensuring high voice quality and low latency. Specific technical challenges addressed include managing conversational chaos, enabling authenticated access to order management systems, and integrating across multiple business systems. The post also discusses legal rules for AI phone calls, focusing on inbound call handling permissibility and outbound call regulations like TCPA. Decagon Voice is presented as an enterprise AI answering platform that combines natural dialogue with authenticated action-taking, featuring real-time responsiveness, optimized generation latency, customizable voice profiles, cross-channel memory, smooth human escalations, SMS integration, and incremental permissions models for authentication depth. It provides an example of enterprise clients like Chime processing over one million calls monthly with high resolution rates.
  45. What AI customer support agents actually do and how they transform support | DecagonThis post elaborates on the capabilities of AI customer service agents, distinguishing them from traditional chatbots. It highlights their ability to perform autonomous execution of complex actions, deep system integrations (e.g., Salesforce, Zendesk, Stripe), and decision-making through reasoning engines governed by Agent Operating Procedures (AOPs). The post categorizes AI agent tasks into identity verification, information retrieval, transactional tasks, proactive assistance, and intelligent routing. It also touches upon voice capabilities and the integration with key business platforms via APIs.
  46. The future of AI agents is test-driven | DecagonThis post introduces Decagon's integrated testing infrastructure for AI agents built on Agent Operating Procedures (AOPs). It details the capabilities of this testing suite, including real-time previews, unit tests for response validation, integration tests for tool triggering and data retrieval, and large-scale simulations. The post emphasizes the importance of test-driven development for ensuring predictability, consistency, and reliability in AI agent behavior, especially when dealing with non-deterministic LLMs. It also highlights the auditable execution paths provided by AOPs for debugging and verification.
  47. Power customer-centric decisions with intent tags | DecagonIntroduces intent tagging as a core component of the AI Support Agent Platform. This feature automatically categorizes customer conversations by their stated goals, providing three levels of granularity for analysis. Intent tags are paired with performance metrics (deflection rates, CSAT, escalation patterns) to identify areas where Agent Operating Procedures (AOPs) may be underperforming. This enables data-driven updates to AOPs to improve customer experience and agent effectiveness.
  48. Why Decagon built Agent Operating Procedures | DecagonIntroduced Agent Operating Procedures (AOPs) as a new framework for building AI agents, moving away from traditional decision trees. AOPs are designed with three core principles: speed to build (natural language for business users, code for integrations), flexibility (merging conversational instructions with code-based logic), and guardrails (code-based validation for sensitive actions). This approach aims to accelerate innovation and improve the adaptability and reliability of AI customer experience.
  49. Simulations: How AI agents get tested and trusted | DecagonIntroduces 'Simulations' as a new testing capability for AI agents, focusing on generating conversations from mock users to test agent logic, knowledge retrieval, and guardrails. This includes specific features for testing AOP selection/execution, knowledge base retrieval, and guardrails/brand tone. It also highlights the use of historical transcripts to create grounded test cases and the ability to simulate voice-specific conditions (accents, speed, noise, tone). The 'Trace View' for debugging decision paths and the 'Watchtower' for automated QA of test runs are also introduced as part of this testing framework.
  50. Watchtower: Always-on QA for every conversation | DecagonIntroduced Watchtower, an always-on QA system for AI agents and human interactions. Watchtower analyzes every conversation in real-time against customizable criteria to identify compliance risks, negative sentiment, or upsell signals. It features natural language flagging, flexible time windows, precision filtering, custom categorization of flagged conversations, and dashboards for trend analysis and drilldowns. This system aims to provide comprehensive, scalable, and context-aware conversational oversight, addressing the limitations of reactive and operationally heavy traditional QA methods.
  51. AI Agents for Fintech Customer Service | DecagonThis post details how Decagon's AI agents are being applied to specific financial workflows within the AI Support Agent Platform. It highlights the integration of fine-tuned LLMs with real-time data access to handle complex tasks autonomously, such as account opening, personal information changes, card management, and personalized product advice. The post also touches on insights and analytics derived from customer interactions, demonstrating the platform's capability to drive higher satisfaction and deflection rates in the fintech sector.
  52. AI agents are never done: The new build-vs-buy calculus | DecagonThis post reframes the build-vs-buy calculus for AI agents, emphasizing that they are living systems requiring continuous engineering investment beyond initial production deployment. It details the hidden costs of building in-house, including ongoing engineering time for workflow updates, RAG pipelines, evaluation, safety, latency, and monitoring; compounding performance gaps; operational risks like on-call rotations and incident response; and the expanding infrastructure footprint for GPU orchestration, inference, vector databases, and observability. The post advocates for retaining control over differentiating layers (workflows, domain logic, brand voice) while offloading non-strategic infrastructure tasks. It critiques service-heavy vendor models and promotes a product-first approach with features like Agent Operating Procedures (AOPs), simulations, agent versioning, Watchtower, and Diagnostic Tools to enable rapid iteration and continuous improvement.
  53. Engineering fast, performant AI agents through fine-tuning | DecagonThis 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.
  54. Introducing Agent Versioning: engineering-grade governance for AI agents | DecagonThis post introduces Agent Versioning, a new capability for the AI Support Agent Platform. It enables engineering-grade governance for AI agents by allowing teams to experiment safely in isolated Workspaces, version and validate changes with Git-based tracking and in-console controls, and govern deployments with enterprise-grade auditability. It also details how data-backed experimentation allows comparison of metrics between production versions and how this integrates with existing GitHub pipelines for structured reviews and deployment workflows.
  55. Converting stress into relief with AI agents for travel and hospitality | DecagonThis post details how Decagon's AI agents are being applied to the travel and hospitality industry to improve customer experience. It highlights five key use cases: handling disruptions and seasonal surges, resolving bookings and account needs, acting as a personal travel concierge, converting service moments into revenue moments, and turning support into operational intelligence. The post references existing Decagon capabilities like Agent Operating Procedures (AOPs) and guardrails for reliable AI agents, as well as Watchtower for QA and Ask AI for querying interactions, indicating the application of these systems to a specific industry.
  56. UndatedIntroducing Duet Apprentice: AI that learns like humans do | DecagonIntroduced Duet Apprentice, a system that enables AI agents to learn from diverse enterprise knowledge sources (onboarding docs, wikis) and real-time team communications (Slack, Teams). This extends the agent's learning beyond production data by analyzing escalated conversations and observing ongoing team discussions, allowing for proactive adaptation to policy changes and edge cases. The system aims to ground agent reasoning in specific business context and improve self-improvement mechanisms.
  57. UndatedPersonal agents are here. Meet them on your terms. | DecagonThis post introduces Decagon's Personal Agent Gateway and PACT protocol. The Gateway provides personal agent detection signals, a dedicated channel with separate Agent Operating Procedures (AOPs) for routing and policy enforcement, and scope-based access control. PACT, built on Agent2Agent and OAuth 2.0, offers a secure authorization mechanism for personal agents to act on behalf of users with defined permissions.