Short answerA long-horizon AI agent runs a task over days or weeks instead of a single conversation — it waits for the right time, wakes on an event, and resumes where it left off. That only works when four mechanisms are in place: a wait with a wake time, a wake cause, a repeating schedule, and an approval floor it can't cross on its own.
In depth
A long-horizon AI agent is an agent that owns a task across days or weeks, not a single conversation. A normal chat assistant answers one turn at a time; the moment the conversation ends, it forgets everything. A long-horizon agent has to survive the hard part, which is not the work — it is the waiting. A task that should take 30 days is rarely 30 days of continuous work. It is a few hours of work scattered across 30 days of silence, with the agent waking at the right moment each time, remembering where it left off, and not doing anything irreversible while nobody is watching. An LLM is stateless, so none of this comes for free. Every time an agent wakes up, it starts from zero unless something durable carries its state forward. Tycoon's agent solves this with four mechanisms that were built and shipped together. Each one answers a different way the task could die. **1. A wait with a wake time.** The agent can park a task until a specific moment — "check again in 14 days" — and go quiet. The task is not running and not consuming resources while it waits; it is scheduled to resume at a named time. This is what turns a 30-day task from a fiction into a plan: the agent does not have to stay awake for a month, it only has to be correct when it wakes. Tycoon keeps every planned task bound to at least one wake source, so a delegated outcome can never silently sleep forever. **2. A wake cause.** A time is not always the right trigger. Sometimes the agent should wake when something happens: a prospect replies, a payment clears, a competitor ships, a metric crosses a threshold. The agent subscribes to the event and wakes when it fires. Tycoon pairs every event wake with a clock fallback — a deadline that still fires if the event never comes — so a task cannot hang forever waiting on a signal that never arrives. That is the difference between waiting and getting stuck. **3. A schedule.** A 30-day task usually needs several beats, not one — check on day 1, day 7, day 14, day 30. A recurring schedule runs on its own cadence and advances one step each time the agent wakes, so the work keeps picking back up on the rhythm it committed to without a human nudging it forward. The schedule is what makes the beats happen. **4. An approval floor.** An agent that runs for weeks will eventually reach a step it should not take on its own — spending money, publishing a public post, committing a launch. The approval floor is a hard gate: the agent keeps working on safe lanes, but it cannot cross the gated step until a human approves. And if it keeps hitting the same wall, it stops and raises a visible blocker instead of retrying forever. This is what makes "leave it running" safe rather than reckless. Underneath all four sits recovery. A long-running task will crash — a process dies, a script fails, a network drops. Tycoon marks a crashed attempt as a crash and writes no verdict for it; the next wake resumes from the interrupted attempt rather than redoing or losing work. No two attempts on the same task overlap, so a retry cannot double-run an action. A deterministic test harness simulates weeks of a task's life in milliseconds and asserts these facts against a durable record, so the behavior is verified rather than assumed. The result is a specific outcome: you delegate a result in natural language and leave. The agent keeps one accountable outcome moving through setup, execution, recovery, and reporting, and interrupts you only when something actually needs you. "Wait 30 days" stops being a promise an agent cannot keep and becomes an ordinary job it just does.
Examples
- An outreach agent that parks a follow-up for 14 days and replies the moment the prospect responds — not when you remember to ask
- A weekly reporting agent that wakes every Monday, pulls fresh numbers, and posts the brief even while you are away
- A launch agent that waits until the build is green, then asks for approval before publishing the announcement
- An agent that parks a task for 30 days and wakes to re-check a metric instead of looping every minute
- A crash mid-task: the agent resumes from its last completed step rather than restarting from zero
- A competitor-monitoring agent that wakes when a rival ships, with a daily fallback in case nothing happens
- A task that hits a spend threshold, stops, and raises a visible request for a human to approve the next step