Fabrika
Case study · AI agents · Platform · 2026

PlutoBuilding the control plane for agents that actually work.

A multi-surface platform for provisioning, operating, and governing AI agents that use tools, delegate work, and run beyond a single chat window.

Pluto homepage with an illustrated night sky and AI agent product message
Overview

Pluto began with a more demanding idea than adding an assistant to a dashboard: what would it take to make an AI agent a dependable member of a team? The answer required product surfaces for deploying, instructing, observing, and governing the work—not only a chat interface.

We built Pluto as a control plane backed by isolated agent runtimes. Each deployment can carry its own tools, skills, memory, credentials, routines, and operational state while the platform keeps people in control of consequential actions.

Product
Pluto
Year
2026
Scope
Agent architecture · Product design · Full-stack development · Infrastructure
The challenge

Agents become difficult to operate as soon as they leave the demo path. Provisioning can fail halfway through, credentials need boundaries, long-running work needs a ledger, side effects need approval, and delegated agents need to return useful outcomes to the parent conversation.

The approach
01

Separate control plane from agent runtime

A SvelteKit and Convex dashboard owns policy, state, provisioning, and audit history. Isolated Eve runtimes execute agent sessions and tools, with signed communication back to the platform and a repair path when configuration drifts.

02

Encode how an agent is allowed to act

A dispatcher classifies work into four lanes: answer directly, delegate to an isolated subagent, draft a side effect for human approval, or stop when a capability is missing. Runs, events, approvals, and idempotency make that policy visible and recoverable.

03

Build for work that crosses surfaces

Web, admin, desktop, mobile, docs, inference, and messaging services share one product model. Routines, connectors, memory providers, computer use, and Mission Control let agents keep working where the team needs them.

What shipped
  • Self-serve agent deployment, provisioning, repair, and teardown
  • Streaming chat sessions and delegated agent runs
  • Mission Control with run state, events, outcomes, and retries
  • Human approval queues for consequential side effects
  • Scheduled routines, connectors, memory providers, and tools
  • Desktop, mobile, iMessage, inference, admin, and documentation surfaces
In production
8

Specialized apps and services in one platform

4 lanes

Direct, delegated, drafted, or blocked execution

24/7

An agent designed to work alongside the team

Under the hood
SvelteKitConvexVercel EveVercelElectronExpoEffectPostgresPhoton SpectrumAI Gateway
Up next
SaaS · Developer platform · 2026

PassBuddyMaking Apple Wallet infrastructure feel like a form builder.

Get started

Ready to build the real thing?

Book a 20-minute call and we'll map the fastest honest path from your current product to something ready for production.