Customizing Agent Behavior

UX / UI Design Visual Design 2025
Customizing Agent Behavior

Overview

Blockit is designed to act like a personal scheduling assistant — handling the back-and-forth of coordinating meetings automatically by understanding your calendar, preferences, and constraints. But scheduling is highly contextual. Users want the agent to behave differently depending on the situation: investor calls versus internal syncs, protecting deep-work time, prioritizing urgent requests. The challenge was giving users fine-grained control over the agent's behavior without turning the product into a complex rule-builder.

Founding Product Designer Blockit AI, San Francisco 2025

Impact

6% → 10% of all bookings Template-based bookings doubled, directly contributing to a Q4 company goal
555 templates, 342 users From 4 templates in Oct 2024 to 130/month by Jan 2026
5% → 41% memory context Agent customization became a core part of how users interact with Blockit
The Problem 01/03
The agent was one-size-fits-all — and users knew it

Blockit managed meetings according to a universal set of preferences — availability windows, meeting lengths, calendar priorities. These defaults worked well for routine scheduling, but users quickly hit their limits. Not all meetings carry the same weight, and they needed a way to tell the agent that.

Blockit had tried to solve this before with codewords and templates, but adoption was low. Most users didn't fully understand how they worked, and the tools felt disconnected from how people actually think about their meetings.

What users actually needed

Rather than starting from the existing features, I focused on what users were really asking for:

  • Trust that the agent behaves correctly in specific scenarios — investor calls, recruiting chats, and internal planning sessions all carry different expectations.
  • A way to signal priority differences — not all meetings should be treated equally. Users needed to express when something should override normal preferences.
  • Control without configuration overhead — they didn't want to manage a rules engine. They wanted to set intent once and trust the agent to follow it.

I also advocated internally for this project to be on the roadmap — it wasn't a given — and owned it end to end from strategy and research through design and execution.

The Approach 02/03
Four principles for AI-native customization
  1. Control without complexity — Users should be able to influence the agent's behavior without building complicated automation rules.
  2. Design around real scenarios — Customization should reflect how people actually think about meetings, not abstract configuration settings.
  3. Meet users where they are — Users shouldn't have to preconfigure every possible scenario. Preferences should evolve naturally as they use the product.
  4. Keep the agent in the loop — Configuration should enhance the assistant's behavior, not require users to constantly manage it.
Scenario-based templates

The solution was to introduce scenario-based templates — reusable configurations that let users define how the agent should behave for a given type of meeting. Instead of abstract settings, users describe contexts they recognize: an investor call, a recruiting screen, a weekly sync.

Each template captures the behaviors they want for that context — priority level, meeting length, links to share, follow-up behavior. The agent then applies the right template automatically based on how a meeting is described.

Grounding customization in real scenarios made the system feel approachable without reducing the control it offered.

The Outcome 03/03
Templates took hold — and kept growing

Template-based bookings increased from ~6% to ~10% of all Blockit meetings within the first two months of launch — directly contributing to a Q4 company goal. The percentage of users with multiple templates grew by 3% in the same period, indicating users were actively building out their configuration rather than trying it once and stopping.

Creation has continued to accelerate well beyond launch. Templates went from 4 created in October 2024 to 130 in a single month by January 2026 — a slow burn that turned into a clear inflection. As of March 2026, there are 555 templates across 342 unique users, and template-present bookings have reached ~10% of all confirmed meetings.

The memory system — launched separately in September 2025 — reinforced this. Memory context went from appearing in 5% of bookings to 41% in seven months, showing that the broader agent customization layer has become a core part of how users interact with Blockit.

What this unlocked

Beyond the numbers, this project gave the agent a mechanism it previously lacked — the ability to behave intelligently across different meeting contexts. That's foundational for Blockit's long-term vision of a truly autonomous scheduling agent.

It also demonstrated something important internally: that thoughtful design of the configuration layer can move a core product metric, not just improve usability. That argument has shaped how the team thinks about future agent-facing features.

Key Decisions

Scenarios over settings

The existing codewords and templates were abstract configuration tools — and adoption was low. Instead of improving those mechanics, I reframed customization around real meeting scenarios users already recognize: investor calls, recruiting screens, weekly syncs. This made the system immediately legible without reducing the control it offered.

Advocating for the roadmap, not just the design

This project wasn't a given — I advocated internally for it to be prioritized. The bet was that thoughtful configuration design could move a core product metric, not just improve usability. The results validated that argument and shaped how the team thinks about future agent features.

Next Project Blockit Onboarding →