Case study · A Cloudstudio original
A platform of AI workers holding real roles — 24/7 shifts, live traces, scored performance and a salary measured in cents.
( The headcount problem )
Every studio carries roles that are pure routine: post the AI news, triage the inbox, draft the proposal, audit the SEO. Hiring humans for them is expensive; dropping them is worse. So we built Agents — and hired software instead.
Each AI worker is a LangGraph workflow with a job title: a schedule, tools, feeds and guardrails. A Python engine executes the graph; a Laravel dashboard employs it — dispatching runs, streaming node-level progress over websockets, and filing a scored report after every shift. First staffed with Cloudstudio's own roles, then with a client's.
The whole payroll costs less than a coffee.
( The personnel files )
The overview reads like a company directory: clients on the left, their AI workers nested under them — a CEO, an Executive Assistant, a Social Media Manager. Run activity, a status donut and 474 scored runs on the books. The engine's heartbeat sits in the corner: online, always.
The workflow library is the platform's job board: Blog Auto-Writer, Email Triager, Executive Assistant, SEO Auditor, Proposal Generator — each a published template of one to ten nodes, tagged by department and reusable across clients. Staffing a new role means configuring a workflow, not writing code.
Every worker gets a personnel page: total runs, success rate, average shift length and — the line that reframes everything — total cost. The Social Media Manager has clocked 206 runs at 94% success for $2.14, running three configured agents with their own feeds, tone settings and editorial thresholds. The run history below scores every shift.
Open any run and the trace draws the workflow on a dotted canvas: fetch the news feed, pick the story, read the page, draft the tweet, post to X — five nodes, each timed to the millisecond, each stamped with a green check as the engine reports back live over websockets. Not a log replay; the actual shift, as it happens.
The report card holds the worker accountable: an overall score, exact AI usage — model, calls, tokens in and out, latency, cost to four decimals — findings if anything went wrong, and a human-readable summary of what was done and why, down to the tweet it actually published. Exportable as markdown, filed forever.
( Under the hood )
A stateless Python engine runs each role as a LangGraph workflow, dispatched over HTTP. Every node reports back through webhooks and Laravel rebroadcasts over websockets — the trace you watch is live, not a replay.
Every run is judged and scored, every LLM call metered to the token and the fourth decimal of a dollar. Reports export to markdown; failures surface as findings. 474 runs on the books, every score in plain sight.
Clients, workers and workflow templates are first-class objects with schedules, actions and API access. The same platform staffs Cloudstudio's own roles and a client's — a new worker is configuration, not a rebuild.
We build AI workers that hold real roles — scheduled, scored and audited down to the cent. This platform staffs our studio; the next one could staff yours.