We reworked how our AI support agent does things. The new approach lets the model take a real agentic action.

AI Agents

As ReleaseDock grows we intend to add more agentic capabilities, particularly, customizable agentic capabilities, allowing ReleaseDock to do things like trigger refunds, upgrades or even book demos.

What changed

Handoff is now a tool the model calls. When the agent decides a conversation needs a human, it invokes a structured handoff tool with a reason and a summary. The tool's handler runs the same deterministic, tested Convex mutation that marks the conversation as escalated. The decision lives with the model; the effect lives in code we can test.

Because the model genuinely understands the request, typos just work. Paraphrases just work. Other languages just work. There's no phrase list to maintain and no second layer of regex trying to reverse-engineer what the model "meant" from its own output.

Why it matters

An AI support agent that promises a human handoff and then silently fails to deliver one is worse than no agent at all — it breaks the single promise the customer cares about most. Routing the decision through a tool call closes that gap structurally: if the model decides to hand off, it calls the tool, and the handoff happens.

Building for what's next

This is infrastructure, not a one-off. Triage, tagging, routing, and suggested replies are all built as tools on the same shared agent, so adding a capability means adding a tool — not another pass of string matching over generated text. One agent, many tools, every decision made by the model and every effect handled by deterministic code.