Essarion Agents.

A project-centric AI workspace where agents read files, run real shells, drive a real browser, and produce auditable artifacts. Not a chat box. A workspace.

§ 01A workspace, not a chat

Most AI products are a single text input wrapped around a single model. The agent answers, the conversation scrolls, and when the tab closes the work is gone. That shape is fine for casual questions. It collapses the moment the work has to produce something — a contract summary you can email, a financial brief you can attach, a spreadsheet of extracted line items, a deck of meeting notes.

Essarion Agents is built around the opposite shape. Each user gets a set of projects. Each project is a persistent container with its own files, its own terminal, its own browser, and its own history. The agent operates inside that container — reading, writing, running commands, navigating the web — and the output of every step is captured, not lost.

You sit on one side of a chat. The agent sits on the other side. Between you is a workspace with four real surfaces it can act through, and a drive of files both of you can see.

§ 02What's in the box

Open a project and you'll see a familiar workspace, with one new thing in the middle.

• Chat. The conversation pane. Where you ask, and the agent answers in plain language.
• Drive. A per-project file area. Drop in PDFs, spreadsheets, contracts, decks. The agent reads them.
• Terminal. A real PTY the agent can run shell commands in — and that you can watch live.
• Browser. A real cloud Chromium. The agent navigates, clicks, fills forms, extracts. You see the page.
• IDE. A read-only file tree and viewer over the project, so you can see what the agent is touching.
• Preview. A live-served view of any dev server the agent spins up — Vite, Next, Flask, CRA.
• History. Every run is persisted. Re-open it, replay it, audit it.
• Jobs. Pre-built typed templates — accountancy, contract summary, executive brief, RFP response — that turn a one-line ask into a structured deliverable.

§ 03What makes it different

Three things, all of which compound.

Real surfaces, not stubs

The terminal is a real node-pty shell streamed to the browser over WebSocket. The browser is a real Chromium running in a managed cloud pool, controlled through Stagehand's natural-language interface. The preview is a real proxy in front of a real dev server. The agent can do things — install a dependency, run a build, log into a portal, scrape a paginated list — and you can watch it happen rather than read about it after the fact.

Persistent project state

A project is not a thread. Files survive. Terminal scrollback survives. Run history survives. You can come back to a deal room, an account, a contract review three weeks later and pick up exactly where you left off — every artifact the agent produced is still in the project, every file you uploaded is still in the drive, every past run is replayable.

Observability of every step

An agent run is an ordered timeline of typed events: tool calls, tool results, terminal output, browser sessions, surface switches, artifact creations. Each event is persisted and exposed both in the UI and as Server-Sent Events. There is no opaque "the agent thought for a while." There is a transcript of what it did, in the order it did it, with the inputs and outputs of each step.

§ 04Who it's for

Three audiences end up in this workspace most often.

• Analyst & ops teams who do the kind of work that has to leave the chat — invoice extraction, transaction categorization, contract summarization, compliance checks. The agent reads the file, produces the deliverable, files the deliverable.
• Executive workflows — diligence memos, board packs, deal underwriting, RFP responses, SOWs — where the output has to be defensible and the trail of how it was produced has to exist.
• Product & engineering teams who want a code-aware agent surface with a real terminal, a real browser, and a real preview. Spin up a sandbox, install dependencies, run tests, ship a draft.

§ 05Where to go next

From here, three paths.

• I want to do my first run end-to-end → Quickstart.
• I want to understand how projects, runs, and sessions compose → Project workspace.
• I want to see the surfaces the agent acts through → Surfaces.