From intent to outcome.
Five stages. Zero hand-holding.

Tell Prometheus what you need — in plain language. "Research competitor pricing and put it in a spreadsheet." "Check my email every morning and summarize anything from the team." "Deploy a three-agent team to audit this codebase."

Requests arrive via web UI, Telegram, API call, or scheduled cron trigger. The system treats every input channel identically.

Prometheus analyzes the request against its available tools, current context, and persistent memory. It builds an execution plan — what tools to call, in what order, and what success looks like.

For complex tasks, the planner may deploy a team of specialized subagents, each assigned a scoped piece of the work.

The plan becomes action. Browser sessions open. Files get written. APIs get called. Desktop applications get controlled. Every tool runs locally with native system access — no sandboxes, no simulations.

Tool execution is recursive: the output of one tool can trigger the selection of the next. The system adapts mid-execution when results differ from expectations.

After execution, Prometheus checks its work. Did the file actually get created? Did the browser reach the expected state? Did the API return the right data? Synthesis rounds catch vague or incomplete results.

For scheduled jobs, a second synthesis pass runs automatically when initial results are ambiguous, producing a clean final summary.

Results, decisions, and learned context get written to persistent memory. The next time you ask about the same project, Prometheus already knows the history. Sessions end — understanding doesn't.

Memory is encrypted, local, and browsable. You control what persists and what gets forgotten.