The quiet unlock of scheduled AI agents

Most automation has a quiet prerequisite: somewhere, a server is running. A cron box, a VPS, a hosted platform that stores your credentials and acts on your behalf. For years, automated trading implicitly meant trusting someone else's server. That is starting to change. Agent runtimes like Claude Code now support scheduled sessions: your AI agent wakes on a timer, does a defined job on your machine, writes down what it did, and stops. No server to rent, patch, or trust.

It sounds like a small feature. It is not. Scheduling is the difference between an assistant you have to remember to prompt and software that simply runs. And trading is unusually well suited to it, because the market hands you the schedule.

What agent scheduling actually is

A scheduled agent session is a full agent run kicked off by a clock instead of a person. At the appointed time the agent starts, reads its instructions, connects to its tools over MCP (the Model Context Protocol, an open standard for agent-tool connections), does the work, and leaves a transcript you can read afterward.

That last part separates it from a classic cron script. A script that hits an unexpected condition either crashes or plows ahead. An agent can notice that the data looks stale, that the broker reported a position it did not expect, or that a number fails a sanity check, and it can stop and say so in plain language. The failure mode of good agent scheduling is a readable note explaining why nothing happened. The failure mode of a cron script is silence.

Markets are a schedule-shaped problem

Trading work does not arrive randomly. It arrives at the same times every day, which is exactly what a cadence is for:

• Pre-market scan. Before the open: refresh prices, re-score the universe, publish today's ranked list, and flag anything that changed overnight.
• Entry checks. During the session: is the pullback you were watching actually holding its level, or did support break? A timed check answers that without anyone staring at a chart.
• End-of-day review. After the close: reconcile what filled against what was planned, log every decision, and note what tomorrow's scan should account for.

None of these steps needs a human at the keyboard. All of them used to need a hosted service, because a human at a keyboard was the only alternative to a server. Scheduled sessions dissolve that trade-off: the cadence runs on hardware you own.

Why local-first beats a hosted bot for custody

A hosted trading bot has one architecture available to it: you hand your broker API keys to the vendor, the vendor stores them on its servers, and the vendor's code trades with them. Every promise about security is a promise about someone else's infrastructure, and you cannot audit any of it.

A locally scheduled agent inverts that. Your credentials never leave your machine. The broker connection belongs to your agent, not to a vendor's backend, and since Robinhood launched agentic trading in May 2026 that connection can be a first-party one. The code that runs on the timer sits in a folder you can open and read. If the vendor disappears tomorrow, your system keeps working, because the vendor was never in the execution path. We wrote more about this model in self-custody trading software and in the Claude plus Robinhood guide.

The honest caveat: local-first shifts responsibility to you. Your machine has to be on, your agent has to be configured, and there is no support team watching a fleet. For some people a managed service is the right call. But if custody of your credentials matters, the local model is the only one where custody is real rather than promised.

What to demand from anything that trades on a timer

The moment software can act without you present, its defaults become the product. Three things are worth insisting on before you schedule anything against a real account.

Disarmed by default. The software should install in a state where it can scan, score, and report, but cannot place an order. Arming should be a deliberate human step, not a config flag that happened to ship as true.

A kill switch you have rehearsed. One command that stops all trading. Run it once before you ever arm, so you know it works and know where it lives.

Logs you can actually read. Every scheduled run should leave a plain-language record: what it saw, what it decided, what it did, and why. If the audit trail is a dashboard summary rather than the agent's own transcript, you are trusting, not verifying.

How Coil runs

Coil (coil.trade) is built on this model. It is a scanner that scores every S&P 500, Nasdaq-100, and Macro-book name on leadership and entry quality, a dashboard that shows the ranking, and an engine that trades the published scores by rule: buy leaders pulling back to real support, never chase, hold cash when nothing qualifies. It runs as scheduled sessions inside your own Claude Code agent, against your own Robinhood account over MCP, with keys that stay on your machine. It ships disarmed, arming is a deliberate step, and every run is logged. The research behind the ranking is laid out at /how-it-works, and if you are weighing this against hosted alternatives, Coil vs trading bots and signal services is the honest version of that conversation.

The pattern is bigger than any one product. Software that runs on a cadence inside an agent you control is quietly becoming the default architecture for personal automation, a shift we cover in agent-native software. Trading just happens to be the domain where the stakes make the architecture's virtues obvious.

FAQ

Does my computer need to be on for scheduled agent sessions?

Yes. Local-first scheduling runs on your machine, so if the laptop is asleep at scan time, that session does not run. Well-built software treats every session as a fresh start and reconciles against the broker's actual positions before acting, so a missed run degrades into a skipped check rather than a stale, dangerous assumption.

Is a scheduled AI agent the same as a trading bot?

They overlap, but they differ where it matters. A hosted bot runs on the vendor's servers with your API keys stored there. A scheduled agent session runs locally, inside a runtime like Claude Code, with credentials that never leave your machine. Both are automation, both can lose money, and both deserve the same scrutiny about what they do while you are not watching.

Can a scheduled agent place trades while I sleep?

If you arm it, yes. That is exactly why disarmed defaults, hard caps, and a rehearsed kill switch are non-negotiable. Coil ships disarmed, nothing trades until a human deliberately turns it on, and it can be shut off in one step. No amount of automation removes market risk.