Multi-Step Web Tasks

What Makes Multi-Step Tasks Hard

Simple tasks (fill one form, extract one page) are where CUAs are most reliable. Complexity compounds errors: each step that can fail creates exponential failure paths in a 10-step workflow.

The discipline of multi-step CUA design is minimizing failure points and building recovery logic for each one.

Task Decomposition

Break complex tasks into explicit phases:

Task: Onboard a new vendor in our procurement system

Phase 1: Vendor research

• Verify the vendor's website and contact information
• Confirm they have the required certifications

Phase 2: System entry

• Log into the procurement portal
• Create a new vendor record
• Fill all required fields

Phase 3: Document upload

• Navigate to the vendor's documents section
• Upload the W9 form (provided as attachment)
• Upload the insurance certificate

Phase 4: Approval routing

• Submit the vendor for approval
• Confirm the approval request was received
• Note the approval request ID for tracking

Each phase is a checkpoint. If Phase 2 fails, the agent hasn't wasted time on Phases 3 and 4.

State Tracking

For long workflows, instruct the agent to save state:

"After completing each phase, output a status update in this format: Phase [N]: [COMPLETE/FAILED] Summary: [what was done or what failed] Data collected: [any IDs, URLs, or data gathered]"

Parallel vs Sequential Steps

Some steps can be parallelized (if you have multiple agent instances):

Sequential requirement: must log in before accessing the portal.

Parallel opportunity: research three vendors simultaneously in three agent instances, then enter them sequentially.

Handling Timeouts

"If a page takes more than 30 seconds to load, take a screenshot, note the URL and action attempted, and try refreshing once. If it still doesn't load after refresh, mark this step as failed and continue."