Legacy vs v4: What you need to know

We’ve revamped Pylon’s support agent to include more cutting edge AI features and give it the flexibility and connectivity to the rest of Pylon’s Agentic Platform. The agent can investigate more on its own, draws on more of your tools and processes, and writes answers straight from your help content.

Summary of Core Changes

The legacy agent answered questions by following a fixed sequence of steps. This gave the agent reliable determinism - several smaller checks decided up front whether a ticket be processed by a runbook or be escalated straight to your team. The agent then searched your knowledge base and separate steps wrote and reviewed the reply. That kept it predictable, but it also limited what it could do.

The new agent works more like an experienced support rep. It reads the conversation, works out what it needs to know, looks it up, and keeps going until it can answer or knows it should hand off. It can search the KB multiple times by executing different searches if previous attempts don’t find relevant content, and use the tools and skills you've set up to execute workflows.

Highlights

• Guidance informs agent behavior: Pylon's Guidances are the general principles users write for the agent to help inform user behavior. Guidances can be used to define a custom tone, format, and persistence when responding to messages, so your agent can behave as your support rep would.

• Skills let you teach the agent your business workflows: A skill is a set of natural language instructions for one kind of request, like a refund policy or a troubleshooting flow. The agent picks up a skill when its description matches the ticket. The model interprets the language in the skill to execute the workflow described.

• Escalation: The agent decides when to escalate based on user-stated guidelines alongside Pylon's own guidelines. A first-class escalation flow be configured by the user and is used to ensure desired escalation behavior once the model decides to escalate.

• Self-healing Agent: The agent learns from tickets answered across the org. A Knowledge agent generates or updates Knowledge Articles based on closed tickets and adds them to an internal KB for agent access, so the agent’s knowledge resources stay up-to-date with each closed ticket. Public facing articles are never generated without user approval.

• MCP Connectivity: Connect custom tools like account and ticket lookups, or apps your team already uses such as Notion and Linear, and the agent can dig into account-specific questions that previously required a human support rep.

• Testing: Improvements to Backtesting, Quality Assurance, the Agent Ticket Playground, QA Mode, and Skill Testing allow users to build confidence in the agent prior to enabling it for live tickets.

What's different from before

The previous agent had built-in rules for when not to escalate, such as when your docs already explain the behavior or when a single clarifying question would do. The new agent doesn't apply those on its own. If your instructions are broad (for example, "escalate anything that looks like a bug"), you may see more handoffs than before until you add those rules to your instructions.

Skills are applied whenever their description matches. A description that's too general can pull a skill into tickets it wasn't written for. If a skill requires a step the agent can't complete, such as checking a tool that's unavailable, the agent will hand the ticket to your team rather than guess.

First replies can take a few seconds longer, because the agent decides what to look up before it searches. We're working on bringing that back down.

Runbooks are no longer supported in the Support Agent.