Decagon Launches Missions for Long-Running AI Customer Workflows

Decagon introduced Missions on November 12, 2026, a product designed to let AI agents manage customer outcomes spanning multiple conversations, systems and participants. The agents can pursue a defined objective for days, weeks or months rather than ending their work with a single interaction.

Decagon cites prior authorization for an MRI, small-loan approval and lead qualification as examples of multi-step journeys that can generate meaningful revenue. For customer-experience teams, the intended change is to replace manual status tracking and handoff coordination with ongoing refinement of the journey.

Orchestration Across Conversations and Back-Office Systems

A team starts a Mission by stating the goal in plain language, naming the personas the agent should engage and adding guidance. The agent then chooses subsequent actions around that objective. It can connect threads involving different people, contact each participant through an appropriate channel and wait when a third party must complete a step before the process can continue.

The product also extends beyond conversations. Through Browser Actions, an agent can take on back-office work such as completing forms. Decagon’s prescription-refill example begins with a patient notification, continues with provider outreach when authorization is required, and includes running claims again on a payer portal. The process concludes when the agent confirms pickup details with the patient.

That example illustrates the product’s broader enterprise role: a single service outcome may require customer communication, provider coordination and payer-system activity. Missions assigns those activities to one goal-directed workflow instead of treating each interaction as a separate job.

Cases and Memory Shape Subsequent Actions

Each Mission case maintains a record of progress, including completed steps and new information that could alter the plan. The agent uses the case to select its next action, arrange proactive outreach and preserve context for later work. Before acting, it checks both the case and customer Memory again to determine whether the existing plan still fits.

Memory supplies preferences and relevant context from other conversations. Decagon’s examples include respecting a customer’s request not to receive calls after 3 p.m. and accounting for a parallel interaction in which the same customer is onboarding a new product while troubleshooting an unrelated issue.

The same recheck applies when a customer changes the expected sequence. If an agent has qualified a lead and arranged a follow-up, but the customer independently schedules that follow-up by email, the agent adjusts its course. The design therefore separates two kinds of context: the Mission case tracks progress toward the business result, while Memory keeps individual preferences and concurrent interactions visible.

Optimization Addresses Long-Workflow Attribution

Decagon identifies attribution as a central difficulty when automated work stretches across many interactions. If a prescription refill is delayed, teams may need to determine whether the problem originated in the initial outreach, the provider handoff or insurance verification. That makes it difficult to isolate the interaction or handoff that needs to change.

Every Mission is built as an Agent Operating Procedure, or AOP. Teams can write, edit and refine that logic, inspect the agent’s decisions and test individual parts of the configuration to identify what needs improvement. Duet and Duet Autopilot provide recommendations aimed at the full customer journey, extending optimization beyond isolated interactions.

The launch consequently links execution and diagnosis in one product model: agents carry an objective across participants and systems, while cases, Memory and AOP-based testing give teams tools to review and revise the path taken. For enterprises considering longer-running automation, the open question is whether those controls can reliably expose weak handoffs and improve outcomes as the number of interactions grows.