stewArda

An intent-aware AI workflow steward that helps users turn messy conversations into observable systems, actionable workflows, and trustworthy automation.

Comment

StewArda

Inspiration

Modern AI assistants are powerful, but they often behave like black boxes. You say something once, automation happens later, and when things change, users are left wondering: Why did this happen? Can I stop it? Can I change it?

StewArda was inspired by a simple question:

What if an AI behaved more like a careful steward—one you could talk to, reason with, and guide over time?

We wanted to design an assistant that doesn’t just respond to prompts, but holds a conversation across time—one that respects intent, explains itself, and invites users to stay involved, especially for workflows that run repeatedly or touch personal data.

What it does

StewArda is a conversational AI steward that helps users turn ongoing responsibilities—like managing school emails, monitoring messages, or summarizing updates—into clear, reviewable workflows.

Instead of jumping straight into automation, StewArda works through dialogue:

  • It talks with users to clarify intent, not just extract commands
  • It surfaces signals and summaries conversationally, without acting prematurely
  • It proposes workflows in plain language, inviting review and approval
  • It adapts over time through explicit tuning conversations, not silent changes

At any point, users can ask:

  • What is this doing right now?
  • Why was this created?
  • Can we change this?

StewArda always responds in context, making the system feel understandable and approachable rather than opaque or fragile.

How we built it

StewArda is designed around the idea that conversation is the interface, and structure is what keeps that conversation safe over time.

Core principles include:

  • Structured artifacts that capture understanding (designs, signals, summaries, proposals)
  • A clear lifecycle separating intent, planning, approval, and execution
  • Human-in-the-loop checkpoints so nothing runs without consent
  • A lightweight agent layer focused on planning and reasoning, not execution
  • A separate execution layer that runs only when requirements and approvals are satisfied

From a UX perspective, users never “fall out” of the conversation. The interface continuously reflects the current state and offers meaningful next steps, while the assistant explains what’s happening and why.

We also built a fully deterministic mock environment so the entire experience can be tested and demonstrated without external dependencies.

Challenges we ran into

The hardest problem wasn’t generating AI outputs—it was preventing confusion over time.

Key challenges included:

  • Avoiding accidental automation from casual language
  • Preventing context drift across long-running conversations
  • Making workflows editable without breaking history
  • Keeping the UI intuitive without hiding important state
  • Teaching the system when to say, “I’m not sure—let’s clarify.”

We addressed these by prioritizing explicitness over cleverness and designing for long-term use, not one-off prompts.

Accomplishments that we’re proud of

  • A system that never silently changes behavior
  • Clear separation between understanding, proposing, and acting
  • Workflows that can pause, resume, and evolve through conversation
  • A UI that always offers meaningful next steps without forcing action
  • An assistant that remains understandable weeks after setup

Most importantly, StewArda feels calm and predictable—users always know where they are and what they can do next.

What we learned

We learned that:

  • Trust comes from predictability, not automation
  • Users are more comfortable reviewing decisions than undoing mistakes
  • Long-running assistants need structure more than clever prompts
  • Asking clarifying questions is often better than acting
  • Systems that can explain themselves age better

Designing for stewardship fundamentally changes how AI feels to use—it becomes collaborative rather than controlling.

What’s next for StewArda

Next, we plan to expand StewArda beyond individual conversations into shared and collaborative environments.

Upcoming directions include:

  • Shared workflows and spaces for teams and families
  • Collaborative approvals and visibility across stakeholders
  • Multi-channel interaction (chat, voice, notifications)
  • Mobile-first experiences for monitoring and quick decisions
  • Deeper tuning and refinement conversations for long-running workflows
  • Broader domain support across communication and knowledge sources

Long-term, StewArda is designed to become a trusted conversational layer between people and increasingly autonomous systems—helping users stay informed, aligned, and in control.

Built With

Share this project:

Updates