Oppizi · 2026 · In progress
Making Soleon readable: an AI agent platform redesigned for the people who build agents
Soleon is Oppizi’s internal platform where agent builders create safe, company-wide AI agents: agents that take work off employees’ plates and raise the quality of what they do.
Its first version was generated quickly, and it showed: every screen tried to show everything at once. Since May 2026 I’ve been redesigning it with a product manager and five to six engineers. The redesign isn’t released yet; this is the work so far.
- Role
- AI Product Designer
- Timeline
- May 2026 – now
- Team
- 1 PM, 5–6 engineers
- Status
- Internal tool, redesign not yet released
What I did
- Redesign of the agent list, the create and edit flow, tools, knowledge, monitoring and admin
- New ways to understand an agent: the agent map and the knowledge graph
- Prototyped and built in code with AI, with Figma where it helps
Problem
The first version showed everything at once
Soleon’s first screens came out of generated defaults, and the quality wasn’t there. It was worst where it mattered most: creating and editing an agent, where builders faced a mountain of controls with no order to them.
Title, description, model, a preview and the editor all sat on one screen, the editor squeezed into one part of it, and a copilot panel stayed open whether you needed it or not.
Options and decisions
Common things first, power when you ask for it
Agent builders are engineers, so I designed Soleon the way tools engineers already trust are designed, like Vercel or Supabase: dense and fast, with nothing hidden that a power user needs.
By default the agent list shows what people scan for most: status, owner, last deploy, the model, and each agent’s skills and tools. Anyone who wants more can customise the view.
One editor, switched instead of stacked
The editor is the part builders spend their time in, so it now gets the screen. Editing and previewing switch in place instead of sharing it, the copilot opens when you ask for it, and everything else an agent has (skills, tools, memory, subagents, automations, evals) moved into tabs along the top.
Settings that rarely change, like the description, stepped back so the agent’s identity reads as one document.
A map, because agents are hard to understand
An agent is a model plus identity, skills, tools, knowledge, automations and evals, and none of the old screens showed how those fit together. Understanding an agent meant clicking through tab after tab.
The agent map puts the agent in the middle and everything it depends on around it, each part with its own health, so a builder sees what an agent is made of, and what’s broken, at a glance.
The full flow
From a chat to production
A builder describes the agent in Soleon’s chat and shapes it in the editor: its identity, tools, skills, configuration, model, automations and evals. Then it goes through staging, and only then to production.
Every step gives the builder control over the agent, so the flow’s job is to keep that control from turning back into clutter.
- 1Describe it in the chat
- 2Identity
- 3Tools and skills
- 4Model and configuration
- 5Automations
- 6Evals
- 7Staging
- 8Production
Craft
Knowledge you can see
The same idea for what an agent knows: knowledge bases became a graph of pages and sources, so a team can see what an agent knows, how it connects, and what’s missing.
One system across the platform
Tools, integrations, MCP Studio, monitoring, users and the audit log share one layout, one table and one set of controls, so learning one screen teaches the rest.
How I work: AI for most of it, my eye for all of it
AI is in about 80–90% of my workflow. I think out loud with it, research how the best tools solve the same problem, and build several working prototypes in code instead of one mockup, then deploy the one that holds up.
Figma comes in where it helps: sometimes at the start to explore, sometimes in the middle to tighten a screen. AI speeds up the making; what ships is my call.