NIKITA
All work
Convergence → Salesforce · Founding Design Engineer & Lead Frontend · 2024 — 2025

AI agents that navigate real digital workflows

Prototype to Salesforce acquisition in ~a year. I owned the frontend and the agentic UX for an AI that perceives, reasons, and acts across dynamic interfaces.

Agentic UXDesign SystemMulti-agentNext.js

Context

Convergence built advanced AI agents that perform complex, human-like tasks in digital environments — navigating dynamic UIs and multi-step web workflows. I joined as the founding design engineer and frontend lead.

Problem

Agentic behavior is powerful but opaque. The hard part wasn't the model — it was making an agent that perceives, reasons, and adapts legible and controllable, so people could trust it with real work.

Approach

I built the frontend from the first prototype to an enterprise product — the architecture, the design system and the standards the team wrote to — on Next.js, TypeScript, Redux Toolkit, Tailwind and Shadcn.

The part I cared about most was the agentic UX. An agent that reasons, calls tools and casts other agents is only trustworthy if you can watch it — so the interface showed the reasoning as it happened, the tools as they were invoked and the agents as they were spawned, as explainable flows wired to the backend's own events, with the controls to step in. With the backend and research teams I brought model output, retrieval results and speech into the UI live.

Under that sat a component library with design tokens, motion primitives and accessibility standards shared across platforms, and the resilience — latency, error recovery, reconnection — that kept the agent feeling responsive. I also shaped the human–AI collaboration paradigm behind the roadmap.

The year, in broad strokes

From the first prototype to the acquisition — about four hundred of my changes in the product — in four acts.

  1. Autumn 2024 · The shell

    The app's frame and navigation, sign-in, the chat and its prompt. And the first way to watch the agent work: a timeline of what its browser saw, fed live by the agent's own events.

  2. Late autumn · Around the chat

    Sessions people could rename, favourite and share; feedback, a scratchpad, settings, a profile; a guide that appeared where people got stuck.

  3. Winter · The business layer, and the foundation

    Dashboards for people, teams and companies; automations with schedules and runs; a hub of prompt templates whose forms turned a blank prompt into a task. Underneath: the typography, the themes (dark became the default), a component library the team could build from, a structure the codebase could grow in.

  4. Spring 2025 · The multi-agent moment

    DeepWork — its guide and toggle, every agent's screenshots side by side, the Flow view. Saved credentials and payments. The public welcome page and the marketing site: the hero video, the stats, the launch posts.

Under the hood

The piece of that agentic UX I'd show first. In March 2025 Proxy shipped DeepWork: one request became a plan, the plan became agents — browser agents reading sites in parallel, a chat agent comparing what they found, a report agent writing it up — and the chat showed one agent's screenshots at a time. People couldn't tell how many agents there were, which one was working, which had stopped at a login and was waiting for them, which had failed and been retried.

So I designed one screen with two lenses on the same run. The Flow view: the orchestrator on top, agents in layers by what they wait on, links that carry work while an agent uses another's results, every state in one vocabulary — cold while an agent works, warm when it needs you or finished, red when it failed — and new agents entering where they will stay, so the picture never reshuffles under your eyes.

Proxy's View: what the agents saw — one window per browser agent with its latest screenshots streamed as it worked, the orchestrator's plan as a log, the comparison and the report as they were written; a tap on a window made it the whole panel with a step back through everything that agent had seen, and when an agent stopped at a login its window said so and the person could take over in a live browser. One clock under both: a tap on an agent in either lens switched the session to it. People watched the work instead of waiting for it — and could step in where it needed them.

Watch a run

A reconstructed DeepWork run — an angel's due diligence on a startup — replayed, with three ways to see the same moment: Flow — the agents and what connects them; Windows — what every agent sees, drawn, not captured; Log — every moment in the agents' words. Five scouts read the site, the launch, the thread, the funding history and the team; one stops at a login until you sign in for it, one fails and is recast, one earns a deep dive cast mid-run; then the analyst weighs the claims and the writer writes the memo. One clock under all three lenses: pick an agent in one and it is picked in the others. Improvised draws a run the script didn't plan.

RunningProxy DeepWork · diligence0:00 / 0:15

Request · Due diligence on Orrery Labs — the warehouse-robot startup from my inbox: the traction, the team, what the market says.

one run · three ways to see it
Legend
Planned
cast, waits on others
Working
at work
Needs you
stopped for you
Done
finished
Failed
gave up here
Waiting
nothing to pass on yet
Ready
results on the table
Flowing
being worked on
Handed
the hand-over is done

Proxy · WorkingReading the request — five scouts, a comparison and a memo.

Runa reconstructed session
OrchestratorWorkingfollowing the run

Proxy

reads the request, casts the others, hands the result back · Read the request, cast the agents, hand the memo back.

Reading the request — five scouts, a comparison and a memo.

Outcome

Convergence was acquired by Salesforce in June 2025.

Key outcomes

  • Prototype → Salesforce acquisition in ~1 year
  • One run, two lenses: a live graph of agents, states and hand-overs, and every agent's browser on one screen with a live hand-over at a login
  • The business layer around the chat: dashboards, automations with runs and schedules, a hub of prompt templates
  • Owned FE architecture + design system, set team engineering standards
  • Shipped explainable agent UX on the web
  • Unified component library: design tokens, motion primitives, accessibility
  • Helped hire and grow the founding engineering team

How Proxy landed

What people posted on X when Proxy 1.0 and DeepWork launched — the public reaction the team read every morning, kept as it was.

11 more posts