Hila Akzin
Résumé
All work

ASAPP · 2024–25

AI agent tooling

From engineer-only configuration to a platform for building, testing, and deploying AI agents.

Role
Staff Product Designer
Timeline
2024–2025
Team
Product and engineering
Scope
Internal tooling and customer-facing configuration

Overview

ASAPP was moving from assisting human agents to automating customer conversations. Configuring AI agents required engineers to edit code. I led product design for the platform that opened configuration, testing, and deployment to internal teams and enterprise customers, defining the product model and core workflows with product and engineering.

Where it started

One file, mixed responsibilities

Intent, business rules, security, and API implementation lived together in a YAML file. Authors struggled to distinguish what controlled the agent’s behavior from what supported its execution.

A task’s YAML file, annotated: available APIs, API instructions, general business policies, customer-specific policies, security rules and API implementation

The platform

Configure with clarity. Deploy with confidence.

Task configuration

Tasks describe what the agent handles and how it responds.

Create task dialog: name, task selector description and task message
Create a task
MobileBill task: task selector description, task message, functions and general instructions
Task editor

Functions and APIs

Functions get their own flow for naming, connecting an API, and reviewing its parameters.

Create function: name and a description of how to use it
Name and description
Select API connection: endpoint, request and response parameters
Select API connection

Validation and guardrails

Flag missing information and show which functions depend on outdated API connections.

Task editor with the name field flagged: Name is required
Required-field validation
Functions list with a warning: some functions are using old versions of API connections
Dependency warning

Deployment

One flow to pick environments, add release notes, and review what will be published.

Global deployment: deploy from, deploy to, release notes and resource type
Deployment setup
Global deployment expanded: the tasks and functions selected for publishing
Deployment review

Explore testing and debugging →

Separate behavior from execution

Tasks define what the agent should do; functions handle execution, giving authors a clearer configuration model.

Catch mistakes before deployment

Required-field validation, dependency warnings, and safe-exit prompts help authors identify issues and avoid losing changes.

Review before publishing

A shared deployment flow makes the destination environment and selected resources visible before changes are released.

Outcome

Beyond engineering

The platform enabled PMs, designers, researchers, and sales teams to configure and test agents directly, supporting experimentation and demos. The work extended into customer-facing tooling for the first enterprise customers.